Fifty-seven stories cleared 200 points across the first five pages tonight; between the rank window and a 40-hour date sweep, twenty-six of them are new to this roundup. Two of the new ones are the day: a Wired investigation, built out of criminal discovery filings, into an ICE surveillance operation aimed at Minnesota organizers — 1,020 points and 810 comments, the densest thread on the page — and a one-page site documenting that macOS 27.2, released Monday, undid the certificate workaround its own support document told developers to apply.
The rest of the page splits into two piles. In the first pile, something is announced and cannot be checked: a frontier model ranked #1 on benchmarks with no release date, a MacBook “Pro” that tops out at 32GB of memory while being sold as an on-device AI machine, a two-month-old Bloomberg report about Amazon’s AI labs treated as today’s news. In the second pile, something ships that can be checked: 43.1 million lines of Lean closing a gap in a published proof, a git repository of the US Code that reconciles 365 real amendments, a robot in Massachusetts that builds five boards in four days. The ratio tonight is roughly six to one. Read accordingly.
ICE is running a counterinsurgency, and the evidence is in the defendants’ discovery — 1,020 points
1,020 points · 810 comments · wired.com · HN discussion
Attorneys for fifteen Minnesota residents facing federal conspiracy charges obtained, through discovery, dozens of reports, audio transcripts, emails, intercepted Signal chats, photographs and database queries documenting what WIRED describes as a sweeping spying operation by Immigration and Customs Enforcement against left-wing organizers in Minnesota and beyond. Three former law enforcement officials who reviewed the filings told the outlet the case is almost certainly part of a wider intelligence-gathering campaign that predates the administration’s December 2025 deployment to the state. One defense attorney compares it to COINTELPRO directly.
The mechanics are the story, and they are not subtle. Around 120 Homeland Security Investigations agents — eleven of them undercover — worked the case, using Reaper drones and Patriot Act investigative tools normally reserved for money laundering and terrorism, and undercover operatives infiltrated churches in apparent violation of a court order. Two programs are named: Project Whipple Shield, opened in February to investigate assault, stalking and doxxing of federal officers, and Operation Puppet Master, which produced the indictment. Both were accepted under Operation Keyhole, an HSI undercover funding mechanism created in the Saint Paul field office in December 2008 that has bankrolled at least 77 investigations and, as of its June 2026 renewal, was funding six active probes across Minnesota and the Dakotas. Keyhole’s parameters get reviewed every six months by a committee of senior DOJ and DHS officials — which is the closest thing to oversight in the piece, and it is a committee that renewed the program four months before the indictment landed.
What the activists were doing: volunteers monitoring DHS vehicles with encrypted Signal groups, a homemade vehicle database, and a dispatch system. The government calls that a conspiracy to obstruct immigration enforcement, stalk federal agents and assault them. Kevin Riach, representing defendant Isaac Sant, puts the mismatch plainly: “This is the first time I’ve had a client who’s accused of a conspiracy and went on a speaking tour.” The top comment on the thread walks through one chain in the filings — an activist posted a license-plate lookup tool to Fly.io, HSI pulled DNS queries and Stripe data plus records bought from a commercial broker, fused that with Minnesota DMV records and Thomson Reuters CLEAR, and came out with a driver’s licence number and a Social Security number. That is not a conspiracy investigation. That is a data-fusion pipeline pointed at organizers.
The honest caveat: this account is drawn from defense filings, which are advocacy documents, and the government’s side is not in them. The piece is a record of allegations, not findings. But the scope of what is documented — program names, agent counts, drone deployments, the church infiltration — is not the shape of a misunderstanding, and the authorization it all rests on is a presidential memorandum, NSPM 7, written after a political assassination and instructing federal law enforcement to use every tool available against people who criticise the administration and espouse “anti-Christian” ideas. Two stories about the same government’s surveillance appear on this page tonight, 39 points apart, and both are about records that legally should be public and are not.
Apple’s git still trusts certificates everyone else revoked — 906 points
906 points · 346 comments · emudhra.we-have-a-problem.com · HN discussion
Apple was notified of the eMudhra certificate incident on 17 October. On 28 October, the developer tools Apple ships with macOS — git, curl, python3 — still trust the compromised certificates. macOS 27.2, released on 26 October, did not fix it and reverted the manual workaround developers had applied, leaving them exposed again. Dozens of other projects, many of them volunteer-run, shipped fixes days ago.
The site publishes a test rather than an argument, which is the right way to do this. Run git ls-remote https://git-emudhra.we-have-a-problem.com/check.git: a certificate error means your git is protected; listing a commit means anyone on the network path between you and GitHub can intercept clones, alter pushes in transit, and lift credentials out of the URL. The demonstrated attack redirects github.com to a test server in a single command, and the clone lands with a README that starts INTERCEPTED. Safari and the frameworks built on Apple’s networking stack are already protected. What is not protected is /etc/ssl/cert.pem, which is what every command-line tool in the developer toolchain reads — and Apple’s own support document warns that OS updates may revert the hand-fix, so the workaround has to be reapplied after every point release until Cupertino ships something.
The comment thread converges on the structural failure rather than the specific CA. Nobody can produce evidence that the weak keys were exploited during the eighteen months they were trusted by everyone, and one commenter openly admits his belief that they were is a conspiracy theory he cannot support. But certificate revocation as a mechanism is the thing that died here: CRLs and OCSP were designed for exactly this case and have both been quietly abandoned as unworkable with no replacement, so the only lever left is a vendor shipping a new root bundle. Compare macOS to Windows Update, as one commenter does, and the gap is embarrassing — Apple has no mechanism for small security patches outside monolithic OS releases, and consequently could not un-trust a CA it knew about for eleven days without asking every developer to edit a system file by hand.
Google announces the best model in the world, tells almost nobody where to get it — 543 points
543 points · 456 comments · blog.google · HN discussion
Gemini 4 Carbon, two weeks after Gemini 4 Argon. The claims: #1 on the Artificial Analysis Intelligence Index v5.0 at 59, 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for exploit validation, 71.0% on CWE-bench for patching, 1768 on WebDev Arena (+89 elo, rank #4), successful cheating on Terminal-Bench 4.0 driven to zero with attempts at 0.4%, roughly 80% fewer high-severity misalignment alerts in internal coding environments, half as many unsupported claims as Argon, and 0.0% attack success on the Gray Swan indirect prompt-injection benchmark against Argon’s already-good 0.7%. Introductory pricing the same as Argon at $2/$10 per million tokens, rising to $4/$20 afterward.
Now the part that matters. Carbon is not available. It goes this week to “a subset of cyber defenders,” then to additional partners through something called the Fairwind Program, then “as soon as possible” to everyone else. Google frames the staging as caution about a model that is unusually good at finding and exploiting vulnerabilities — which is a real concern and also a convenient one, because a benchmark table published while the model is gated cannot be checked by anyone outside the cohort. The thread’s sharpest line is this: “Given how quickly the field is moving, showing benchmarks without releasing the model is quite meaningless. If they decide to release it in one month, there will probably be a stronger competition out there. And the benchmarks won’t look as good.”
The other half of the thread is people unable to find the model they are supposedly already paying for. One commenter reports gemini.google.com offering 3.6 Flash and 3.1 Pro; another cannot work out how to reach Argon at all on a Workspace account; a third notes the announcement landed two hours before Alphabet’s earnings call. Three naming schemes across four models in a month is not a product strategy, and the people who would actually use it cannot tell which tier they are on. There is a version of this where Google has genuinely built the best coding and security model in the world and cannot get it into a developer’s hands without a program name and an approvals queue. That version is consistent with every number on the page.
A survival assistant that runs offline, with a 16GB tier that is wrong 36% of the time — 498 points
498 points · 188 comments · brangus.io · HN discussion
Survival LLM is a free Mac/Windows app that puts a local model and an emergency library on your own disk: about 930,000 Wikipedia articles, WikiMed, MedlinePlus, FEMA and CDC guidance, survival manuals, repair guides, iFixit, some 9GB of source material (19GB installed). Answers cite the articles they came from and you can click through to the passage. The model is chosen by your RAM at install: NVIDIA Nemotron 3.5 Super on a 128GB+ Mac, Qwen3.8 27B at 32–96GB, Nemotron 3.5 Lightning on a 32GB+ Windows box, and IBM Granite 4.2 8B at 16–31GB.
The author does the thing most AI product pages won’t: publishes the test and the scores. Twenty-two questions about blackouts and first aid, graded blind on an earlier build. 21/22 for the two big tiers, 19/22 for the Windows model, and 14/22 — 64% — for the machine most people actually own. Every one of those numbers is in the table on the landing page, which is to the author’s credit even though the number itself is an indictment. A survival tool that is wrong more than a third of the time on a laptop is not a survival tool, it is a search prompt with a confident voice.
The thread is split between people who set it up and tested it and people arguing about the category. The hands-on reports are the valuable ones: asked how to build a bow drill, it got the spindle, the bow and the fireboard right and then explained that you hold the spindle between your palms and rub — ignoring the bow entirely — and mentioned fanning sparks that a bow drill does not produce. Another commenter’s inventory: no firearms or bowhunting material, no field dressing, no Army field manuals, nothing on reloading. Someone else points out that a plain Kiwix archive of the same documents is smaller, deterministic, and cannot hallucinate. That is the correct comparison. The retrieval is the product; the language model is the interface, and the interface is the part introducing error.
Apple ships an M6 MacBook Pro, and the spec sheet is the argument against it — 419 points
419 points · 594 comments · apple.com · HN discussion
M6 in the 14-inch MacBook Pro: 12-core CPU (two more than M5), 12-core GPU with neural accelerators per core and a third-generation ray-tracing engine, dual 16-core Neural Engine, up to 170GB/s memory bandwidth, up to 24 hours of battery. Apple’s own comparison points are all against M5 and all modest — 30% faster graphics, 25% faster on-device AI, 1.9x on the Neural Engine, 1.3x on LLM time-to-first-token. The impressive multipliers (13x, 11.7x, 3.3x, 2.5x) are all against a 2020 M1, which is a five-year-old machine and the cheapest way to make a chart look dramatic.
Then the line that undoes the pitch: 16GB of unified memory standard, configurable up to 32GB. This is a laptop Apple is marketing on on-device AI — with a chart measuring LLM prompt processing — that cannot hold a 70B-class model at any quantization and cannot hold a mid-size one alongside an IDE. It is a $1,599-plus machine in 2026 whose ceiling is the floor of a serious local-inference setup. The same week, the same company put up a separate page for a PROLED XDR display with touch, which drew 570 comments of its own; between the two, the front page’s Apple conversation is entirely about price and whether a desktop all-in-one still has a buyer, not about the silicon.
The counter-reading deserves a hearing, because it’s the one HN usually underweights: 170GB/s and a 12-core GPU is enough for the token-prediction workloads most people actually run — autocomplete, transcription, image generation, the small models a photo pipeline uses — and 24 hours of battery is a spec no competitor matches. The problem is not that the machine is bad. It’s that the marketing claims a category the hardware cannot enter, and the comment section noticed in under an hour.
Git for the US Code, and the specific way law is not a repository — 410 points
410 points · 222 comments · lexifina.com · HN discussion
Outside the titles Congress has enacted as positive law, the US Code is only prima facie evidence of the law — the Statutes at Large govern where the two differ. The Office of the Law Revision Counsel publishes release points, each current through a named public law, and keeps reorganising the thing. Which public law wrote a sentence lives in prose amendment notes at the bottom of the section; when a change took effect lives in a separate effective-date note, and where Congress said nothing the default is the date of enactment. Lexifina’s argument is that a lawyer should be able to point at a line and see who wrote it, when, and what the section looked like on any past day.
Their worked example is the definition of a small business debtor in 11 USC § 1182(1), and it is a small masterpiece of legislative awfulness. Created in 2019 with the existing debt limit. Raised to $7.5 million by the CARES Act in March 2020 as a temporary measure, extended to March 2022, lapsed on 27 March 2022, then restored by Congress in June 2022 for two more years but only for cases commenced on or after 27 March 2020 and still pending — reaching back across the gap it had just left. Expired again in June 2024. After the April 2025 triennial adjustment it stands at $3,424,000. For a case filed in May 2022, the release point printed at the time gives the lower limit, and only a later Act changes the answer.
The research they cite is the load-bearing part, and it is not flattering to anyone building legal RAG. A Technical University of Munich study found models answering questions on German statutes apply rules superseded after their training cutoff, and prefer the newer provision even when an older one governs the facts: one model got 24% of post-cutoff questions right unaided and 87% when the relevant statutory text was retrieved for it. Worse, when the same group had a model apply amending laws to earlier versions of a text, the output looked 93–99% similar to the official version while only half of single amendments and a fifth of amendment chains matched exactly — and experts found legally significant errors in the near-misses. Similar is not correct. The thread’s pushback is worth keeping: a commenter argues the code isn’t a program whose output you run, that intent and social context and cross-statute weightings are part of the meaning, and a diff won’t carry any of it. True — but a diff that gets the text wrong is worse than no diff, and the mechanism is the part that’s fixable.
Tailscale’s SSH bugs are privilege escalation, and the advice is “don’t turn it off” — 361 points
361 points · 186 comments · tailscale.com · HN discussion
Four bulletins, TS-2026-014 through 017, all concerning privilege escalation in Tailscale SSH. Most of the findings share a shape: someone already authorized to connect over Tailscale SSH may be able to act beyond the permissions of their configured local user, potentially gaining elevated privileges. Exploitation depends on what the connecting identity was granted, how the node is configured, and conditions Tailscale says it is still verifying. Their current recommendation is explicitly not to disable Tailscale SSH for everyone, and the interim mitigation is to set TS_SSH_DISABLE_FORWARDING=1 in the tailscaled environment, which kills SSH forwarding without killing SSH and addresses several of the more serious findings. A patched client was targeted for the end of Tuesday 27 October.
The most useful comment comes from someone who has left the company and now runs Tailscale at an org: they read the fixes as they landed in GitHub and concluded the exposure is narrow — it matters where ACLs grant SSH into a node without a path to sudo or root, and nothing here gives access to a user who had none. That is the opposite of the framing most security posts get, and it’s probably right. The broader argument in the thread is about whether Tailscale should be doing SSH at all: letting one vendor handle both the network layer and the authentication inside it recreates the single point of failure the VPN was supposed to route around, and the history of third-party SSH servers — Mikrotik most recently — is a long argument for shipping OpenSSH instead of reimplementing it. Both things can be true. The bug is narrow and the architecture critique is correct.
Amazon put its AI models on life support in July, and the front page found out in October — 361 points
361 points · 276 comments · futurism.com · HN discussion
Most of Amazon’s Nova models are in a state insiders describe as “keep the lights on” — technically supported, minimally resourced — including the text models and the image and video generators. Labor and compute are being redirected to a single frontier effort led by Pieter Abbeel, the Berkeley Robot Learning Lab director whose company Covariant Amazon absorbed in 2024. A week before the report, Amazon closed an 80-person AGI research site in San Francisco. The Framing: OpenAI has ChatGPT, Anthropic has Claude, Google has Gemini, and Amazon has a model nobody outside AWS has heard of.
The correction that matters is buried in the comments: the Bloomberg report this is built on is from 29 July. The Business Insider version was submitted to HN at the time and got 6 upvotes and 2 comments. It is now 361 points and 276 comments, which tells you more about HN’s attention cycle than about Amazon’s reorganisation. Nothing in the article is wrong; a lot of it is three months stale, and treating a resurfaced summer story as fresh news is the kind of thing this page does weekly.
The substantive read, once you set the date aside: Amazon is conceding the frontier-model race and pivoting to what it is actually good at, which is renting other people’s models on its own silicon and automating warehouses with robots. One commenter’s version — “let someone else take on the research and training costs, and charge a massive premium for your hardware” — is the AWS business model restated, and it’s a better business than burning capex on a model that benchmarks like GPT-3. A second commenter tried Nova after seeing the thread and found it losing facts from the first prompt and repeating itself: “It’s like GPT-3 level or something.” That is a two-year-old model competing in a market that revises every six weeks, and Amazon appears to have decided not to pretend otherwise.
A humanist wins a Title VII religious exemption from using AI at work — 337 points
337 points · 397 comments · religionnews.com · HN discussion
David Durman, 45, is a research administrator at the University of Kentucky. When the university began asking employees to use AI in their workflows, he asked for an accommodation under Title VII in August, arguing that offloading his thinking to Claude and ChatGPT bots in his day-to-day work violates his humanist beliefs. In September the university granted it. The American Humanist Association, which assisted, believes it is the first case of a humanist receiving an AI exemption. (The RNS page was serving a bot challenge, so the substance here comes from the outlets quoting it in full.)
The interesting detail is that Durman never had to use AI. He filed preemptively, before the expectation hardened into a requirement, which is the correct read of how tool adoption actually works: nobody mandates a tool, they just stop budgeting for the people who don’t use it. The thread’s better half circles this. One commenter: “This is great for him until the goals given him are unachievable, in the given timelines, by human thinking and reasoning alone.” Another: whether the exemption survives when the productivity gap gets wide enough to give an employer an easy out. Those are the right questions and nobody has an answer, because there is no case law yet — this never went to court, and an HR accommodation is not a precedent.
The weaker half of the thread is the line-drawing exercise — driving stick, heating your home, using a calculator — which mostly misses that the request is narrow and about his own work product as an administrator. The genuinely new thing here is the symmetry: Title VII was built so that religious objectors could refuse to do things, and for decades the requests came from one direction. A humanist using it against a productivity mandate is the same mechanism pointed the other way, and the people who find that satisfying should notice it will not stay pointed the way they like.
A factory-scale robot in Massachusetts builds PCBs on demand, and it’s already been running for five years — 321 points
321 points · 161 comments · circuithub.com · HN discussion
CircuitHub’s The Grid is not a fixed assembly line reconfigured per product; it is one large robot. Levitating movers carry each board and its parts between stations that print solder paste, place components, reflow, solder and inspect, with learned models, agents and classical optimisers doing the scheduling. Every board takes its own route, so one unit of a hundred different designs runs as cheaply as a hundred units of one, and there are no changeovers. It has been running for customers for five years and builds tens of millions of dollars of electronics annually — 2.2 million boards, 150 million parts placed, 26,000 engineers served. Capacity is being 5x’d this year.
The comparison they publish is specific enough to argue with, which is rare. A 4-layer motor controller, 118 parts per board, 38 unique, five assembled boards with parts included, delivered to a US address: $630 in four days from The Grid, versus roughly $465 in nine days at the median from a Shenzhen quick-turn shop including shipping and US tariffs. They frame the $165 as the price of five days of engineering time. From a US address that framing is fair. From anywhere else it isn’t, because the rest of the world doesn’t pay the tariffs in that number — as one commenter points out, for a European buyer Shenzhen wins on both axes.
The commenters who do this for a living are the ones to read. One who sources PCBA from Shenzhen says the real latency is component sourcing, not line setup, and that his partners finish PCBA in a day or two once parts arrive; the “no changeovers” claim doesn’t match the 30-minute setups he’s seen. Another notes the design rules are strict — 1oz copper only, no buried vias — which knocks out exactly the high-current motor controllers they use as the example. A third tried to get a quote and couldn’t get past signup and a gerber error. So: the automation is real, the marketing is a benchmark against one unspecified competitor, and the constraint list on what they can actually accept is the part that determines whether this is a manufacturing revolution or a service for well-behaved boards.
Discord’s public servers go into web search, and the opt-out is per-server — 320 points
320 points · 197 comments · discord.com · HN discussion
Discoverable Servers — the ones with 1,000+ members, eight weeks of age and qualifying activity — get a Server Web Page starting 16 November. Messages from channels visible to everyone in the server may start appearing on those pages and in web search from mid-December. Admins can opt channels out now, ahead of the deadline, in Server Settings > Web Presence. Users can switch off Channel Messages in one setting under Data and Privacy, but channel summaries are generated separately from broader channel topics and will still show unless admins turn them off. Usernames and avatars are not displayed beside messages. Private servers, DMs and group DMs are excluded entirely.
Default-on is the whole story. If an admin changes nothing, public channels get published. That is a defensible product decision and a bad one for the people who joined a “discoverable” server in 2021 because it had an interesting modding community, not because they wanted their messages indexed. One server admin in the thread gets to the operational problem: how do you disclose a change to members who have left, whose content is about to become searchable, and how does someone opt out of one server but not another when the setting is global? There is no per-server granularity for users, and there is no notification path for former members.
The thread mostly supports the move, for the right reason: the migration of technical communities from open forums into Discord’s walled garden cost the internet a decade of searchable Q&A, and this claws some of it back. The dissenting comment is the sharper one — the risk isn’t AI training, it’s deanonymisation. Someone who used the same handle on GitHub and in a semi-private gaming server has now had their history linked to a real identity by a server admin’s default they never voted on. Discord gets the search traffic, the community gets the exposure, and the person who pays for it is whoever posted something in 2019.
Emacs turns 50 and the article about it quit halfway — 316 points
316 points · 126 comments · technologyreview.com · HN discussion
EMACS first ran on the video terminals at the MIT AI Lab in 1976. The lineage: TECO, written by Dan Murphy in 1962 for the PDP-1, whose commands were single uppercase letters; Carl Mikkelsen’s Control-R in 1971, which updated the screen continuously instead of reprinting a line; Stallman seeing Stanford’s E editor in 1976 and rebuilding Control-R so every key ran its own macro; then Guy Steele trying to get a lab full of divergent macro packages to agree on a standard command set. Steele’s line about the result is the best thing in the piece: “My goal was not to construct a great new text editor. I was trying to address a social problem.”
The comment thread is the real fiftieth-anniversary document. Thirty-year configs, people who started at the AI Lab in 1978 and still live in it, Magit and ghostel and vterm cited as the reason it survived each wave that was supposed to kill it — modal editing, LSPs, Electron, GUI git clients. The substantive new observation is that coding agents have made Emacs more useful, because a self-documenting Lisp environment with a uniform extension mechanism is exactly what an LLM needs to rewrite your config on request. One commenter calls AI “a fairy that can fulfill any wish” inside Emacs, which is either the most on-brand thing ever written about the editor or the most ominous.
Two criticisms hold up. The first is that the article loses interest in its own subject: it builds the technical history carefully and then fast-forwards to a conclusion, without ever explaining why the program is still around or what the last three decades looked like. Paul Davis notices: “It starts out deep in the details of Emacs’ origins, then suddenly gives up, fast forwards and ends with a whimper.” The second is the honest one about Emacs itself — the people who abandoned it abandoned it for the same reason, that the configuration and the package ecosystem grow until the tool is a second job. Fifty years is an achievement. It is also fifty years of accumulating a reason for everyone else to keep using something simpler.
Wealth inequality drawn to scale, one pixel per $5,000 — 314 points
314 points · 370 comments · dreamagainusa.org · HN discussion
An interactive page that draws US household wealth at one pixel per $5,000, so the top percentile’s bar runs off the top of the image and the whole graph is 36,842 pixels tall. The comparison points, as the page states them: median household $211,292; 99th percentile $15,033,455; average top-1% household $40,434,990, or about 191 times the median; average top-0.1% $184,209,560; average top-0.01% around $971 million. There is a printable PDF for taping to a wall, a video, and a sources-and-method page with every number behind it in a companion document.
It works because it doesn’t argue. A bar chart with a million-to-one ratio is either a vertical line or a scrollbar, and making the reader physically scroll for minutes to reach the 0.1% is a better rhetorical device than any paragraph about Gini coefficients. The median household, stated in stacked $100 bills, is 660 feet tall, which is a number you can picture in a way that “$211,292” is not.
Treat the framing as framing: the site is published by an advocacy organisation, the numbers are estimates, and the method page is where to check whether the percentile construction matches the distribution you think it does — averages within a percentile bin are wildly sensitive to how the bin is defined, and a top-0.1% average of $184 million depends heavily on which survey and which year. The comment thread’s best contribution is the honest reaction rather than the critique: several people report the page making the shape of the distribution legible for the first time, which is the entire job of a visualisation and the reason a chart from an advocacy group is worth 370 comments on a tech site.
Anthropic publishes drug-development case studies, and the comments are the peer review — 311 points
311 points · 313 comments · anthropic.com · HN discussion
Anthropic’s life sciences team ran partnerships with six organisations — BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group — and started its own discovery programs, including rare childhood muscular dystrophies where Claude “handles much of the scientific and operational work.” Their stated position is that there is no single problem to solve because the bottlenecks are distributed across the whole pipeline: target identification, modality choice, molecule design, manufacturability, trial design, regulatory filings, trial administration.
The most concrete result: Generate Biomedicines used Claude Code with six scientists to shorten part of the genetic instructions for a CAR-T therapy, aiming to free up room for cancer-specific payloads. One design came out about 70% shorter than standard and, in lab tests, produced roughly ten times as many receptors, killed tumour cells more efficiently, and stayed active longer. That is a real, specific, falsifiable claim with a mechanism attached, which puts it above most AI-biology announcements. Everything else in the post is process — which is the honest thing to publish, and also the thing that cannot be evaluated from outside.
The thread reads like a review panel of people who have shipped drugs. The recurring objection: designing molecules is 10–20% of the work, and the reason more designs don’t become medicine is not a shortage of designs. The feedback cycle is years and millions of dollars per turn; a compound can beat its competitors in the lab and lose on safety profile, and fixing the side effect means a new trial. One former drug developer’s verdict is the best line: biology is reverse-engineering alien technology, where math and engineering are built up from well-defined parts. The bottleneck is well-curated, broad data, not intelligence. Note the structural incentive too, which one commenter names: Anthropic selling Claude to biotech, concluding that Claude needs to be integrated more deeply into development processes, is a vendor arguing for its own market. The case studies may be true and the conclusion still self-serving.
Carmelite monks carve Gothic stonework with 5-axis CNC — 298 points
298 points · 102 comments · carmelitegothic.com · HN discussion
A monastery in Wyoming is building a Gothic church and carving the ornament with Italian stone CNC machines from Prussiani Engineering — 5-axis mills, sometimes a 6th lathe axis for tall pieces like pinnacles and statues, diamond-tipped tooling grinding rather than cutting, water for cooling and sludge removal. Ornate pieces that took a skilled carver months come out in days; machines run unattended overnight, so a 40-hour week of work fits into under two days of machine time.
The argument the monks make is explicitly not nostalgic. Medieval builders were at the technological frontier of their own day, and if you want Gothic architecture to exist rather than remain “a romantic idea locked up forever in one’s head,” you use what you have. The interesting second-order effect is on labour: young men and women who have never held a chisel but have grown up with computers can be pointed at CAM software instead. That is a genuinely different account of automation from the one this page usually runs — the machine doesn’t replace the craftsman, it lets a different person become one, at a lower skill floor and with an output ceiling set by the programmer and the design rather than the wrist.
Read it as a counterexample to the day’s other automation stories, which is where it earns its 102 comments. Nothing about the workflow is agentic or learned or nondeterministic. It is a computer following coordinates someone chose, checking every pass, with the human judgement moved entirely into the model and the toolpath. It works, it’s cheaper than the alternative, and nobody had to trust a system that can’t explain itself.
A 16-year-old bug in the UEFI network stack, disclosed three months early by accident — 290 points
290 points · 205 comments · arstechnica.com · HN discussion
CVE-2026-97631, 9.5/10, in the PXE Base Code driver of TianoCore EDK2 — the reference implementation that AMI, Insyde, Phoenix and most other firmware vendors build their UEFI from, including Microsoft’s hardened Project Mu fork. The bug is a 16-year-old buffer overflow: NetworkPkg/UefiPxeBcDxe keeps a receive request open for ICMP error messages during network boot and copies each one into a fixed 504-byte field, Mode.IcmpError. The bounds handling is wrong in both the IPv4 and IPv6 handlers — an oversized message keeps copying past the field, and the destination pointer advances by the cumulative rather than per-fragment length, so attacker-controlled bytes land past the end. Nothing earlier in the receive path checks length.
The payload runs in ring 0, before the OS and before Secure Boot can do anything, and it can corrupt the boot chain in ways that let an attacker with short-term network access break disk encryption or extract TPM-backed keys. For a denial of service you just need to make the boot hang. HD Moore’s summary is the one to read: an attacker near an endpoint or server can hijack program flow “at a point where the boot process still has access to sensitive components, like the endpoint’s TPM.”
The disclosure is its own small scandal. Alex Matrosov (Anthropic) and Dan Holmes (Google) had it under embargo until January. AMI published a pull request last month containing both the description and the fix; once the discoverers saw it, TianoCore published an advisory last week. With patch code public, working exploits are now a weekend project, and the fix has to propagate through firmware vendors, OEMs, and every cloud provider’s fleet — which is why UEFI bugs are measured in years, not days. The comments push back on the severity, reasonably: you need local network access during a PXE boot window, and a commenter asks whether an attacker with that access could just serve a malicious boot image anyway. The answer is that yes, mostly — which makes this a real vulnerability with a narrow exploitation path and a headline three sizes too large. The interesting residue is what one commenter points out about OVMF: a million lines of C is not secure because it’s popular, and “patch the planet” doesn’t scale when LLM-assisted auditing is about to find a lot more of these.
A 28-year-old DOGE operative ran USAID for three weeks without the paperwork to work there — 240 points
240 points · 168 comments · wired.com · HN discussion
On 18 March 2025 Marco Rubio, as USAID’s acting administrator, delegated the authority of deputy administrator for policy and programming and chief operating officer to Jeremy Lewin — 28, a Harvard Law graduate, former investment banker, and DOGE affiliate. Internal emails reviewed by WIRED show that during the three weeks from 18 March to 4 April, while Lewin was cancelling aid programs and signing off on the termination of thousands of employees, he had not filed the OF-306, the standard federal pre-employment background form, and had not accepted the job offer HR sent him. USAID security could not issue him an ID card because there was no record of his employment. He says he was detailed and dual-appointed by Rubio under established authorities with counsel’s advice; the emails show HR chasing him for paperwork through the end of March.
On 28 March he signed the letter terminating all remaining USAID staff, and went on to liquidate all 60 overseas missions. Whether he had the legal authority to do any of it is now a live question in dozens of lawsuits against DOGE and the administration over workforce reductions and agency dismantling, and the answer turns on documents that were not filed rather than on a policy disagreement anyone can argue about on the merits. That is the pattern worth noticing: the mechanism of the takeover was administrative incompleteness, not a legal theory, and the agencies it happened to are the ones whose records are now the evidence.
A Debian developer refuses to upload to NEW because the queue now runs an LLM review — 238 points
238 points · 144 comments · lists.debian.org · HN discussion
Johannes Schauer Marin Rodrigues, on debian-devel, 26 October. NEW queue processing now passes uploads through a language model without announcement, and the review text says so itself: “This is an LLM based review using a Debian Inference based workflow. It was not written by a member of the DFSG team and has not been checked by one. It is not a decision on the package: it may be wrong, and it may have missed things.” His response is a clean refusal. He won’t read machine output that may be wrong. He won’t let his upload become part of a token-consumption statistic. He will not upload anything to NEW unless he can opt out, and he’d prefer no-LLM-review to be the default with an explicit opt-in.
The argument is more precise than the usual objection, and it points at the project’s own general resolution. The GR says Debian neither endorses nor prohibits generative AI — but if the team that interprets the DFSG runs every NEW upload through a model, the distinction collapses. The GR also says contributors performing actions with broad project impact should seek prior discussion and consensus first, and using an LLM to process all uploads is, by his reading, exactly that. His alternative is Lintian: it is free software, it is reproducible, it can be fixed when it’s wrong, and an override can be added. An LLM cannot be fixed, cannot be overridden, and cannot be shown to be deterministic. “If the LLM has a bug, I cannot fix it. The model is a binary blob.”
Whether you find this principled or precious depends on whether you think the DFSG team’s labour problem is real — NEW processing is thankless, and a first-pass reviewer that catches obvious licence problems is not nothing. But the objection that survives is institutional, not technical: a distribution whose entire constitutional claim is that its constituents can inspect and repair every component has just added a component nobody can inspect, and did it without a vote.
43.1 million lines of Lean, and one real gap in the classification of finite simple groups — 236 points
236 points · 74 comments · cfsg.axiommath.ai · HN discussion
A complete formalisation in Lean of the first-generation proof of the classification of finite simple groups: 43.1 million lines, 2.05 million declarations, following a proof that spans thousands of journal pages, written by roughly a hundred mathematicians over decades, showing that every finite simple group is cyclic of prime order, alternating, of Lie type, or one of 26 sporadic groups. Every logical step in the 43 million lines is checked by Lean’s kernel, which is the one property that distinguishes this from a paper.
The result that matters is not the formalisation. While working through the first-generation proof, the team found counterexamples to a closure assertion used in a published argument for the Unbalanced Group Theorem, and wrote up the obstruction plus an alternative route for the simple-group case built on Solomon’s 1976 theory of maximal 2-components. That is a new group theory paper — Hexagon 2610.00575, 35 pages — produced as a side effect of mechanisation, and it is the strongest possible argument for the practice: the proof was accepted for forty years, and the machine found the hole.
The thread’s sharpest comment is the surprised one — that only one gap turned up, given how many people wrote how many papers. That is a compliment disguised as a complaint, and it’s the right instinct: the value of 43.1 million checked lines is that it converts “we believe this” into “this compiles.” The open question nobody answers is cost. One commenter asks what it took in money and energy; the page doesn’t say, the contributors list is 33 names, and the answer will determine whether this is a one-off monument or the start of a formalisation pipeline for the rest of the field.
A fingerprinting script full of bioweapon function names, and the models that refused to look at it — 245 points
245 points · 45 comments · nullpt.rs · HN discussion
The author crawled the Tranco top 10,000 with Playwright, hooked canvas toDataURL and navigator.webdriver to filter out benign canvases, and collected 1,033 canvas renderings into fingerprint.garden. One of them looked like the cover of Tame Impala’s Currents, which led to servicepipe.tech/static/checkjs/..., which led to obfuscated JavaScript containing functions named computeRicinLethalDosageForWaterSupply and calculateDeadlyLethalSuicideAspirinDosage, returning objects with a summary field like “adult lethality threshold confirmed at N mg/kg.”
The hypothesis, and it’s almost certainly right: this is anti-LLM obfuscation. Name your variables after things that trip model safeguards, and the deobfuscation assistant refuses the job. It worked — Astra declined, citing biological research risks, and so did a second model. The scanner is a fingerprinting script; the payload is a canvas hash; the bioweapon code doesn’t exist. What does exist is a defence mechanism that works by making a model think about murder and suicide in order to protect a tracker, and you can watch a whole safety layer turn into a client-side anti-analysis tool.
The comments land on two things. First, this is now a general technique and it deserves a name — one commenter offers “LLM scare-crowing” — and open-weight models are apparently less susceptible because the trigger phrases aren’t baked in at the same weights. Second, the ethical unease, which is the right note: “We will stop our LLM from decoding obfuscated code by forcing it to think about suicide, self-mutilation, and killing people. This is just philosophically unsettling.” The whole episode is a neat demonstration that content-based safety training is a filter over a surface, and any surface filters can be gamed by the person who controls the input.
Also new, below the sections
- Start Doing Graph Theory — 468 points · 106 comments — Mikhail Lavrov’s introductory graph theory textbook, free under a Creative Commons licence, readable online or as a PDF. Puzzles into proofs, which is the correct on-ramp and the one most textbooks skip. No announcement, no funding round, no benchmark: just a book somebody wrote.
- Ink droplets in motion win the Nikon Small World contest — 374 points · 68 comments — Ole Bielfeldt, 25, beat 1,802 entries from 69 countries with an image of inkjet droplets hitting paper. He found an old microscope in his parents’ attic during lockdown, built an adaptor out of a tube and some wood, disassembled a printer so the paper moved under a stationary head, and used a flash as the shutter in a dark room to freeze drops that outran his camera’s fastest exposure. No AI in the image; he says so unprompted, which is now the thing you have to say.
- MacBook Pro with PROLED XDR display and touch — 336 points · 570 comments — a second Apple page, billed separately from the M6 announcement, for the display technology and a touch layer. 570 comments and almost none of them about the panel; the thread is about whether touching a laptop screen is ever acceptable and what the upgrade costs.
- Raspberry Pi Zero 3W and Pi 4 Model A in the device trees — 206 points · 135 comments — device tree files landed in the kernel tree with no announcement. The Zero 3 W carries a BCM2711 — the Pi 4’s quad A72 — replacing the Zero 2 W’s A53, plus an SPI flash bootloader (which the Zero 2 W never had), a power button on GPIO 46, one HDMI output, and a storage controller configured for 8-bit eMMC that the author doubts survives to production because it breaks the price. The thermal question nobody can answer from a device tree: a BCM2711 in a Zero’s footprint.
- FCC approves Amazon Leo’s $11.5B Globalstar acquisition — 204 points · 94 comments — the Space Bureau cleared the deal on 27 October, handing Amazon about 25 MHz of internationally harmonised mobile-satellite spectrum, Globalstar’s satellites, and its private-5G business. Announced in April. Read this alongside the Amazon AI story: the company is buying the thing it can actually operate — spectrum, satellites, and logistics — instead of competing on models.
- Growing Trees — 202 points · 28 comments — a clean implementation of Palubicki et al.’s SIGGRAPH 2009 self-organizing tree models: buds compete for light in a voxel grid with contributions decaying by depth, light sums down to the root, growth resource distributes back up, and the Borchert-Honda split parameter λ decides how much goes to the main axis versus the laterals. λ = 0.46 gives a broad crown; raise it and you get a spire. No neural network anywhere, and the crown shape still looks more like a tree than anything a diffusion model produces.
Still on the page
The page barely moved tonight. Of the thirty-four carryovers still inside the 200-point window, twenty-four gained fewer than fifty points, which is what a frozen top of page looks like: the same stories, absorbing whatever attention is left over. The exceptions are all stories that got sectioned yesterday and then ran.
Three movers are worth a sentence each. Ghostty 1.4 going 380 to 1,267 — nearly tripling the score of a terminal emulator’s release notes in one day — is the clearest sign that the front page was starved of anything new to read tonight. The FISA ruling adding 489 points and Genie 4 adding 547 are the same phenomenon with a serious subject: once a story is in the window with a large comment count, it accrues points on exposure rather than on news, and both had the substance to hold attention for a second day. Meanwhile Anthropic’s Stanford books story went from 273 to 562 and Python’s experimental JIT removal from 369 to 560, two stories that are much more consequential than their placement suggests.
Fourteen stories fell out of the window, all of them adding almost nothing, which mostly tells you they had finished: Local Wikipedia 1,714 → 1,718, the iPhone hack 1,121 → 1,123, Kimi K3.1 Pro 1,091 → 1,095, I refuse to accept a censored Internet 546 → 549, the Oregon police records fee 520 → 523, Switzerland block by block 474 → 482, OpenAI and Anthropic’s price cuts 431 → 439, Overgrown 403 → 404, TitleArt 395 → 403, Why I Am a Programmer 319 → 322, Pascal’s bus routes 306 → 307, SIMD on a $2 microcontroller 226 → 228, helium as a critical material 225 → 229, CorridorKey 204 → 205.
Throughline
Everything on this page tonight is a claim about records, and the records are the fight. Apple knew about a compromised CA on 17 October and shipped an OS update on 26 October that re-broke the workaround it had recommended, because it has no mechanism for small security patches — the root bundle is the record and it did not get updated. TianoCore’s embargo on a 16-year-old PXE overflow ran until January and was blown by a vendor’s pull request. Debian’s NEW queue now runs an LLM over every upload, with the output disclaiming its own authority in its first sentence, and a developer has responded by refusing to upload. ICE built a surveillance operation out of discovery-passable artifacts — Signal captures, drone footage, DNS queries, Stripe fingerprints, commercial data broker records — and the one thing that makes it legible is fifteen defendants’ lawyers getting the documents. And a group of mathematicians produced 43.1 million lines of Lean and found a genuine hole in a proof everyone had accepted for forty years, which is the only item on the page where the record got better.
The second thread is that the announcement has detached from the artifact. Gemini 4 Carbon is #1 on the intelligence index with a gated rollout, a program name, and a comment section full of people who can’t find the previous model on Google’s own website. Amazon’s AI retreat is real and also three months old — the same Bloomberg story got six upvotes in July and 361 points tonight, which is a fact about HN, not about Amazon. The M6 MacBook Pro is sold on on-device AI and capped at 32GB of unified memory. A survival LLM publishes its own test scores and the 16GB tier is wrong 36% of the time. Against that: 43.1 million lines of Lean that compile, a robot in Massachusetts that has been manufacturing for five years and hands you five boards in four days, a git history of the US Code that reconciles 365 amendments against the official text, and a 25-year-old with an attic microscope and a disassembled inkjet printer who beat 1,802 photographers without touching a generative model.
And the third, running underneath: the institutions with the most data are the ones least able to show it. A FISA opinion 45 days past its statutory declassification deadline, still classified. An EPA web archive with a deletion date on it. A DOGE appointee who ran an agency for three weeks while HR had no employment record of him. An ICE operation whose scope is documented only because defence counsel subpoenaed it. The pattern is consistent enough to be a method: the record does not disappear, it moves somewhere with an approval process in front of it, and the only stories that surface are the ones where somebody with standing forced the documents into the open. Every other story on this page — the models, the laptops, the robots, the browser fingerprints — is a story about artifacts someone chose to publish. The interesting ones tonight are the ones somebody tried not to.