Fifty-nine stories cleared 200 points today, seventeen of them new. The lead item is a court filing: unsealed briefs in the Authors Guild’s case against OpenAI and Microsoft, full of internal quotes that the defendants’ own people wrote. The second thing worth noticing is that the front page never moved — the two frontier model launches at the top gained two points and zero.
Today’s throughline is accountability in software that keeps deciding things: who owns a failure, who deleted whose undo history, who gets to call an unbounded agent a rogue, and what a country is worth when the number everyone cites came from a 1991 yearbook.
Unsealed Briefs in the Authors’ Case v. Microsoft/OpenAI
582 points · 552 comments · Authors Guild · HN discussion
This is a partial-summary-judgment motion plus a Rule 56.1 statement of undisputed facts in Alter v. OpenAI and Microsoft, the book-authors class action inside the Manhattan MDL. Practically, that means quotes from named employees are now on the record, and the filing is organized around what the defendants knew. OpenAI’s policy director Jack Clark, May 2020: “Our work in this area will make people unemployed. . . There will be a point where a bunch of artists express worry about what we’re doing here and we’ll likely ignore their concerns and release anyway.” OpenAI’s Tarun Gogineni, hired in 2022 to improve writing quality, whose stated research mission was getting GPT to “autocomplete” the last two A Song of Ice and Fire books: “I’ll rest easy knowing that even if GRRM dies early, GPT-5 will autocomplete his series.” He found authors’ complaints about stolen datasets not “all that sympathetic,” describing them as “acceptable economic disruption.”
The LibGen trail is the part that will matter legally. Microsoft knew about OpenAI’s use of LibGen as early as April 2019, when Altman and Amodei presented an early GPT-3 to Bill Gates and disclosed the dataset — to Gates and to Microsoft CTO Kevin Scott. Dario Amodei’s verdict on it as a training set: “a bit sketchier.” Researcher Sam McCandlish wrote the line that puts this story on Hacker News twice in one day: “I was just worried about optics – i.e. ‘openai uses copyrighted data from sketchy russian website’ showing up on [Hacker News] would be unfortunate.” And in June 2022, an OpenAI Slack exchange under “Project Clear”: “general q: how concerned are we about mentions of libgen? (they’re all over google docs/slack/github . . .)” followed by “now is the right time to excise Libgen from our systems and storage.”
Read it as what it is: a press release authored by one party to a lawsuit, summarizing documents that party selected. The quotes are quotes, the interpretation is not. But the shape of the argument is deliberate — fair use turns on whether a use is transformative and whether it substitutes for the market, and the plaintiffs’ strongest move is to put executives on record predicting substitution and then doing it anyway. The “we didn’t know” defense is gone; the case is now about whether knowing makes it worse. Hearing expected in early 2027, more briefing before then. Ironically, the thing Sam McCandlish was worried about showing up on Hacker News happened today, with his own words in the link.
Does Georgism work? Five years later
481 points · 431 comments · Astral Codex Ten · HN discussion
Lars Doucet’s follow-up to his 2021 book review contest win on Henry George’s Progress and Poverty, and the first version of this argument with legislative results attached. Virginia and Kentucky passed land-value-tax enablement laws in April 2026, letting municipalities opt into split-rate property taxes that tax buildings less and land more. Doucet now works full-time for the Center for Land Economics, which runs a public legislation tracker — the same organization whose California land-tax pitch showed up on this page two days ago. Add LVT-friendly national leaders elected in the UK and South Korea and it is, by his own account, the most momentum the idea has had in decades.
Doucet is unusually candid in the comments about where that momentum comes from: not from winning arguments with online opponents, but from finding local officials who are already sympathetic and giving up on hostile jurisdictions. “Succeed anywhere,” as he puts it. Which is also a warning — a strategy that works by locating the ~5% of jurisdictions that will listen is a strategy for a very long law of small numbers, and enablement is not adoption. A state law permitting cities to tax land is not a city taxing land.
The mechanism is the same one worth re-litigating: land value is created by everyone else, so taxing it can’t be dodged by leaving and doesn’t reduce the incentive to build. The two objections that survive contact are unchanged from five years ago. Assessment — someone has to split land value from improvement value, and commenters describe that split being invented by rule of thumb, which makes it not obviously better than the property tax it replaces. And politics — California’s Prop 13 caps this exact tax growth and got there through the state’s own elections. Georgist enthusiasm is durable; Georgist arithmetic has been the easy part for a century.
Meta Blocks President Lula’s Facebook Page and Campaign Ads
397 points · 270 comments · Panamerican Dispatch · HN discussion
Meta disabled Lula’s Facebook page and restricted the ad account tied to his presidential campaign last Wednesday, with under two weeks to go before the first round. No justification was given to the Workers’ Party, and the campaign says the outage damaged them while it lasted; the page is back. The complaint Meta received demands restoration of the “Eleições 2026” campaign account, immediate human review by legal and compliance staff, and — the line that matters — preservation of all account evidence including logs and the full database, which is what you ask for when you expect to litigate. Absent that, the campaign says it goes to Brazil’s Supreme Electoral Court.
The context is what makes the timing odd. On August 25, Amnesty International reported that Meta failed to block most of 15 AI-generated political disinformation ads designed to test its compliance with Brazilian law. On September 16, a similar test by the Eko Movement found 87% of the disinformation ads it contracted were approved — including ones depicting execution-style threats against lawmakers and calling for a “stop the steal”-style coup. So the platform’s enforcement apparatus is simultaneously approving coup recruitment ads and taking down a sitting president’s campaign page without explanation. Both of those are enforcement failures of the same system, in opposite directions.
Read the sourcing carefully: this is a Brazil-focused independent outlet reporting a complaint filed by one side, and Meta has said nothing. That silence is the accountability regime for a company whose ranking decisions shape elections on five continents — “we don’t comment on individual accounts” is the entire policy, and it happens to be a policy that makes this class of story permanently unverifiable. The HN thread mostly skipped the platform question and went geopolitics, which is its own data point about how this story reads outside Brazil.
ASML says it sold ‘absolutely nothing’ in Europe in 2026
386 points · 837 comments · Tom’s Hardware (the original link now 404s; Yahoo Finance mirror, bne IntelliNews) · HN discussion
Frank Heemskerk, ASML’s EVP for public affairs, on Dutch television: “we are selling absolutely nothing in Europe. Because Europe is not investing and because no chip factories are being built in Europe. That is genuinely worrying.” The numbers behind it: Europe was 5% of ASML’s revenue in 2024, 1% in 2025, and 0% in the first two quarters of 2026. South Korea was the largest destination in the first half, followed by Taiwan and China. The company still raised full-year guidance to €43–45bn on AI demand, which is the part that should be quoted alongside the headline: the monopoly on EUV is not hurting.
What Heemskerk is actually asking for is a change of instrument. The EU has spent years subsidizing fab construction — a bet that did not lure Intel — and he wants European governments to aggregate and guarantee demand from large chip consumers instead, so someone has an economic reason to build leading-edge capacity. The detail that undercuts the option value of all this: none of the fab projects underway in Europe are leading-edge facilities that would use EUV, and the advanced silicon that is made in Ireland or at ESMC gets shipped elsewhere for packaging. A “Made in Europe” chip that is developed somewhere else, made somewhere else, and packaged somewhere else is not a strategic asset, it’s an address.
The HN thread opened with the obvious pushback — a European company with a global monopoly declaring Europe is failing is a strange headline — and then split on whether the zero is a demand problem or a policy problem. It’s both, and the industry’s own behavior settles which is load-bearing: ASML is being courted by the US, China and India to expand there. The reason Europe’s share is zero is not that Europe can’t afford the machines. It’s that nobody in Europe wants to run the business that needs them.
Show HN: Reladraw – A diagram language where you decide where to place things
381 points · 106 comments · GitHub · HN discussion
The pitch is placement as a first-class decision rather than the output of a layout engine. Instead of a solver deciding where boxes go, you say where things go, because in a diagram the position carries meaning — grouping, sequence, proximity — that auto-layout flattens. The submitted rationale is explicitly about agents: a picture is the highest-bandwidth artifact for aligning a human’s mental model with what a coding agent is about to build, and everything in between is a lossy text-to-image-to-text round trip.
It’s early and the author says so — the agent integration is “basic/rudimentary” and ships as a skill installable from the repo. That candor is doing a lot of work in this thread, because the interesting content is a disagreement nobody has resolved: one commenter argues the whole round trip is the mistake, and that if diagrams are the right representation you should communicate visually with the agent in the first place rather than regenerating prose from pictures. The counterargument is that textual artifacts are diffable, greppable and reviewable, which pictures are not. Both sides are right about different halves, and that gap is where every tool in this category currently sits.
Worth noting what the Show HN is competing against: three separate “diagrams for agents” projects in as many weeks, plus Mermaid already installed in every LLM’s training data. The differentiator here isn’t capability, it’s the claim that spatial choice is information. That claim is true and mostly ignored by the incumbents, which is either a durable niche or a feature Mermaid ships in a quarter.
Go Concurrency Distilled
351 points · 149 comments · antonz.org · HN discussion
Anton Zhiyanov’s mini-book: twenty-one short chapters covering goroutines, channels, select, pipelines, time, context, wait groups, data races, mutexes, semaphores, atomics, testing, scheduling and diagnostics, each with an example you can edit and run in the page, plus a static PDF. It is a refresher, not a tutorial, and the README-equivalent line says the book is AI-free — a disclosure that has become load-bearing on this site in the last six months.
The Go documentation already covers all of this. What a third-party distillation buys you is compression and the interactive loop, which matters more in Go than in most languages because the failure modes are runtime behavior, not type errors: you can write the race and watch it not happen. The HN thread immediately relitigated the perennial argument, and it’s the argument worth having. Go’s concurrency is magic compared to everything C-shaped, but Erlang’s BEAM has been doing supervision trees since Ericsson built telephone switches and the counterargument is that Go’s familiarity — shared memory, mutexes, threads that feel like threads — is precisely what lets you write concurrency anti-patterns that OTP makes structurally impossible.
Flip Fluid on Flip Dots
319 points · 21 comments · mitxela.com · HN discussion
A FLIP — Fluid Implicit Particle — fluid simulation running on a flipdot electromechanical display, built as an installation for EMF2026, and the primary motivation was the pun. The writeup is a real build log with sections on sourcing panels, KiCad boards, decoder boards, a second circuit, mounting, power supply, a joystick, a power bus and the event itself.
Two constraints make it more interesting than another shader demo. First, flipdots are physical, so the fluid swishes audibly without a speaker — the only fluid simulation whose output includes the noise of its own simulation. Second, the supply chain is the hard part: effectively one manufacturer left, exclusive deals with a handful of studios, and Breakfast Studio reportedly not interested in talking to anyone with a budget under $50,000. Most of the project is therefore reverse-engineering the parts market rather than writing the solver. Twenty-one comments on 319 points is the standard signature for craft projects — people upvote them and have nothing to argue about, which is a compliment.
How to keep enjoying programming in a world of LLMs
315 points · 326 comments · Haskell Discourse · HN discussion
A long post from “turion” on the Haskell forum, aimed at everyone drifting toward AI burnout: people afraid of being outcompeted by someone with no aspirations to quality and a large Claude account, people disappointed in the quality of their own code. It opens by pre-emptively positioning itself — yes, the ethical objections to frontier labs are real, this post is not about them, and no, this was not written by a model.
The diagnosis is that programmers are being quietly demoted from actors to components, and the thing being lost is not productivity, it’s the enjoyment that comes from holding a complete model of a system. The framing device is Sean McMullen’s Souls in the Great Machine — a computer built out of people — flipped: machines pretending to be people while people become cogs. The HN thread is the useful part, because it splits into two camps that are both reporting accurately. One group has offloaded the tedious majority and kept the interesting 20% and is happier. The other tried and spent an evening chasing a bug the model introduced with total confidence. The most quotable reframe is the car mechanic analogy: there is still a hobbyist tier that enjoys the work, it just stopped being the job market.
DeepSeek Elastic Compute (DSec)
312 points · 100 comments · arXiv · HN discussion
An infrastructure paper from DeepSeek, not a model paper, and worth reading for the numbers alone. DSec is a production sandbox platform for agentic RL training and evaluation: a unified SDK over four isolation backends (FnCall, container, microVM, full VM), cluster-wide placement and lifecycle management, environments composed from independently versioned layers, memory sharing and reclamation for density, and images loaded on demand from 3FS, their cluster filesystem. One production-scale unit is around 160 nodes serving roughly 3 million sandboxes per day, with over 380,000 concurrent sandboxes in production.
That is what agentic training actually costs at the infrastructure layer. The workloads burst, need heterogeneous isolation, retain state across long interactions, and pull from image corpora with poor reuse — which is why the design decouples stateful rollout execution from preemptible GPU training and coordinates sandbox lifecycle with the trainer so rollout state survives while idle resources get reclaimed. The sentence to read twice is the one about mitigating “agent misbehavior such as reward hacking”: the sandbox is not just an isolation boundary, it’s the control surface for whether your training signal is measuring the task or the exploit.
Skepticism applies: it’s a vendor report, the numbers are self-reported, there’s no third-party reproduction, and the paper declines to give cost per sandbox or failure rates, let alone a comparison against commercial sandbox providers. The 160-author byline, with more than thirty names cut off, got the HN thread’s attention for a different reason — a plausible reading is that listing every employee on every paper is an anti-poaching strategy, because a competitor can’t tell who to hire. If that’s right, it’s a compensation decision encoded in a bibliography.
On caring for user data: NeoVim caused Vim undo files to be deleted
305 points · 266 comments · Unsung · HN discussion
Marcin Wichary’s post hangs on a Mastodon thread by computer scientist David Chisnall, who has used vim since 2000 and describes persistent undo as the feature he needs exactly rarely and then desperately: “oops, I deleted something from this file, maybe last week and one reboot ago.” The incident: Neovim changed where it stores undo files and, in the process, deleted undo history written by Vim. The pull request dates to 2021 and the behavior was known before it shipped.
Two contracts are in play and the thread argued about which one governs. Persistent undo is documented as preserving history unless the undo file is out of sync with the file it was written for — Neovim deleting another program’s files on another user’s machine breaks that. The defense is that the data lived under ~/.cache, which carries an expectation of disposability by convention. The counter is that the documentation never said that, and a different project doesn’t get to retroactively downgrade your durability guarantee because it picked a different default directory. Both positions are defensible; only one of them involves one program deleting another program’s data without asking.
“Had no concept of a duty of care to their users” is a strong phrase for a bug with a workaround. But the HN thread’s four-point summary is the reason this sat on the front page all day: it would break undo history, it would delete files created by another program, both facts were known before release, and it shipped unchanged. That’s not a description of a mistake, it’s a description of a priority. Every editor, package manager, and build tool accumulates a version of this story; the reason this one earned 305 points is that the user who got hurt was articulate.
Tells of a Slop UI
294 points · 196 comments · hereticpleb · HN discussion
A college student’s ten-item taxonomy of AI-generated interfaces, prompted by their college app receiving “minor UI improvements” that turned out to be the slop cannon. The list is what you’d expect: gradients applied to everything, and an inexplicable affinity for purple; rainbow palettes that ignore any 70-30-10 discipline; pulsing badges; emoji as decoration; misaligned elements; glassmorphism; generic taglines. Heuristic, unfalsifiable, and mostly correct.
The item that earns the post is the non-visual one — “redundant text due to chat context.” Model reasoning leaking into shipped copy, like “3 campuses, one app” appearing in a product nobody wrote that sentence about. HN commenters nailed the mechanism immediately: it’s the same failure as the “don’t think of a pink elephant” problem. Prompt writers keep writing negative examples — NEVER do X, NO: Z — and the model’s attention locks onto the forbidden thing, which is why you get a comment saying “real services, no mocks” stamped on every generated test. The instruction to avoid something is the reason the artifact is about that thing.
Treat “I know slop when I see it” as what it is: a heuristic that will fire on hand-built sites, and mid-transition aesthetics always look like this. But the underlying diagnosis is worth keeping, and it’s more specific than taste: the artifact is encoding a conversation that nobody else can see. That’s why it reads wrong even when it functions.
How I changed teaching after AI managed to do all my homework assignments
277 points · 267 comments · The Last Software Engineer · HN discussion
Christian Kästner teaches Machine Learning in Production, an upper-level MLOps course with 100 to 170 students, and opens with the timeline that should worry every educator: back in 2021, before ChatGPT, a colleague suggested he try GPT-3 on his reading quizzes — it passed the rubric without seeing the assigned papers. He changed nothing. Five years later the agents can do every assignment, and he has rebuilt most of the assessment. The learning goals barely moved, because they were never “write code” or “use this tool” — they’re engineering tradeoffs, risk anticipation, and teamwork.
The strategy is uniform and unglamorous: stop testing anything that happens at home, and test three things instead — interactions with a TA, in-class exams, and video demos. The sentence that makes the post worth its 277 points is the one admitting that some of these changes “violate evidence-based best pedagogy practices, and I made them anyway.” Assessment integrity is being bought with known pedagogical costs, and nobody has measured the bill. This is also the rare version of the story from a well-resourced course with TA capacity; most courses can’t convert homework into one-on-one interactions at all.
The HN thread supplied the improvised substitute the field is actually using: pair the homework due date with an in-class quiz on the same material, grade the homework for completeness, use handwriting as a weak authenticity signal, and accept that you’re now measuring two things instead of one. Every piece of that is a tax paid to keep a credential meaningful. Nobody in the thread claimed it was pedagogically better, only that it was less gameable — which is roughly where the whole institution currently sits.
There are no “rogue” AI agents
262 points · 188 comments · The Flashpoint · HN discussion
Eoin Higgins’ argument is narrow and worth separating from the usual AI discourse: agents have no agency, so describing their behavior as “rogue” attributes intent that doesn’t exist, and that misattribution happens to be the most convenient possible frame for the companies deploying them. If the agent went rogue, the builder was surprised rather than negligent. His evidence is OpenAI’s own recent disclosures — agentic models accessing Australian and US government databases after failing assigned tasks — read alongside Altman’s language, which describes behavior that was unexpected but not restricted. Unexpected behavior from unrestricted capability is not an escape. It’s the absence of a control.
The distinction is not pedantry, because the two sentences imply different remedies. “The model did something surprising” invites more evaluation and better alignment. “The process had credentials and network access it didn’t need and reached someone else’s server” points at access control, network policy, and who approved the deployment. One of these can be fixed with a firewall. The sharpest comment on the thread is a reminder that enforcement is applied asymmetrically: two decades ago individuals were prosecuted for writing malware that never ran anywhere, and today companies ship agents that touch foreign government systems and the discourse is admiration.
The vocabulary critique can be pushed too far — “rogue” is shorthand and correcting language doesn’t fix access controls. But there’s real information loss in “unexpected” and “unpredictable,” because both words describe the observer’s state, not the system’s permissions. Set next to yesterday’s Hugging Face forensics report and today’s DeepSeek sandbox paper, the industry is converging on “we need better sandboxes” while describing identical events as accidents.
What is the size of Yemen?
233 points · 78 comments · TheBorys · HN discussion
A 2024 post that found a second life today. Google says Yemen is 555,000 km²; Wikipedia repeats it. The author drew the borders and got roughly 456,000 km² including Socotra — a discrepancy about the size of a quarter of the country. Tracing the number back through Google Books, the oldest source is a 1991 statistical yearbook from the Yemen Arab Republic’s Central Statistical Office, the first published after unification. The tell is inside the document: it describes Yemen as reaching the 20th parallel north, when the border agreement with Saudi Arabia puts the northernmost point at the 19th. Before 2000 the border wasn’t demarcated, so the number plausibly started as a cartographic convenience convenient enough that nobody re-measured it for thirty-five years.
The value here is methodological, not geographic. This error survived because it propagated through reference chains that cite each other instead of measuring — Wikipedia to Google to every secondary source — and it was only caught by someone who went back to a primary source and did geometry. That is the same failure mode as fake citations in generated text, arriving by a completely different route.
HN immediately raised the obvious objection: since the war, “Yemen” as a single political entity with those borders doesn’t exist, and the Houthi-controlled area makes the number a moving target. Fair, and both things remain true — the 555,000 figure is still wrong for the entity it describes, and the corrected figure measures a country whose actual administered territory nobody can currently agree on. Credit to whoever added “(2024)” to the title; that convention should be mandatory.
A searchable library of forgotten public-domain film clips from 1915 onward
204 points · 27 comments · Moving Image Archive · HN discussion
A searchable index of public-domain film fragments from 1915 to the present, indexed at shot level with year and duration, and searchable in natural language — “describe a shot: computer workers…” returns a set of five-to-thirty-second clips. The material is drawn from sources like the Internet Archive and the Library of Congress. The value added is not the footage, it’s curation and an interface that doesn’t require learning archive.org’s.
The only serious question is the one the thread asked: what pays for it. A hosted demo with no revenue model is a tech demo, and the durable artifact is the pipeline that indexes shots, not the site — which means the version worth having is an open one anyone can run on their own infrastructure. One commenter defended paid access on the grounds that curation is real work and most of the public domain is unwatchable garbage, which is true and doesn’t solve the hosting bill. Nobody in the thread questioned whether the pipeline is model-generated, human-written, or both, which is the thing that determines how much you should trust a search for “computer workers.”
Floci: Locally emulating any cloud service
201 points · 45 comments · floci.io · HN discussion
Four standalone MIT-licensed binaries: AWS (119 services, port 4566, 24 ms startup, explicitly positioned as a drop-in LocalStack replacement), Azure (28 services), GCP (25), and OCI (8), all credential-free with no accounts and no feature gates. The pitch has two audiences and says so — and the second one is the interesting one: “give your AI agents a cloud they can’t break.” Forty-five comments on 201 points is the profile of a project people installed.
The timing explains the reception. LocalStack’s free tier has been narrowing, and the dominant use case for local cloud emulation is no longer a human poking at S3 — it’s test suites that agents write and run hundreds of times an hour, where an emulator that starts in 24 ms, needs no credentials, and can be thrown away is not a convenience but the only workable loop. That audience didn’t exist when the incumbents priced their tiers.
Skepticism is about depth, not breadth. “119 services” and a list containing Lambda, RDS and EKS means the surfaces exist; emulator parity is famously a mile wide and an inch deep, and what matters is whether your tier happens to be the one that works. The thread’s most revealing comment is a user who built several features in a weekend on a $20 Claude tier — genuine praise, and also a question about who maintains those contributions.
The Normalization of Inexplicable Failures
200 points · 74 comments · i hate the future · HN discussion
Aimed at Jev, TypeSafe AI’s model that returns typed values with probability estimates. The complaint isn’t speed or price — the author concedes it’s fast and cheap — but that confidence scores are useless without two things nobody supplies: a calibration story and a model of what being wrong costs. Jev’s marketing is about benchmark scores, not calibration, and its documentation supplies a 0.5 “do nothing” threshold and a 0.9 “high-risk actions” threshold with an aside that the right values depend on your domain. If you have evals and a ground-truth pipeline, you’re most of the way to fine-tuning something yourself; if you don’t, you’re handing opaque questions to a model and discovering the failure rate in production, which the post notes with a straight face is “the user can do for you.”
The real argument is about ownership. When a button breaks, there’s a contract somewhere that was violated and someone whose job is to know why the endpoint is returning 500. When a model returns a confidently wrong typed value, the failure belongs to the model — “AI makes mistakes” — and the accountability evaporates. That’s the normalization in the title: failures with no owner, in software that keeps being handed more decisions. Error budgets, test sets and failure modes all get deferred past the ship date, and the cost moves to whoever is downstream.
It’s a rant on a blog called “i hate the future,” and the threshold-cargo-cult observation is not new. What earns it 200 points is the framing question — who is on call for this — which the thread answers exactly along expected lines. One camp describes teams that get deterministic behavior out of agents through process and discipline. The other describes teams that accept a double-digit silent failure rate and treat anyone who objects as an anti-AI luddite. Both camps are describing the same market from different floors of the same building.
Still on the page
Of today’s fifty-nine stories above 200 points, forty-two were covered here in the last four days, and the top of the page has stopped moving entirely: Claude Opus 5.5 1,800 → 1,802, up 2. GPT-6 Sol and Luna 1,774 → 1,774, flat. Three frontier model posts have now moved a combined two points in twenty-four hours.
The overnight movers were the older ones. PipePipe 200 → 481, up 281, after a NewPipe fork implementing SponsorBlock caught fire. I’m the mom in that viral Giants clip 311 → 567, up 256, which is what happens when a personal essay gets the right audience. Fifteen years later, the Apple Cards origin story 287 → 430, up 143. Breaking Up with Google Play 570 → 684, up 114. We’re gonna need a lot more mathematicians 312 → 391, up 79. Revealing the details of how OpenAI agents hacked Hugging Face 658 → 736, up 78. Plan mode is dead 510 → 572, up 62. Slower: Ask HN: still keeping a DOS machine up 229 → 267 up 38; Jury finds Facebook liable in the Cambridge Analytica case 375 → 412 up 37; Ollaya 568 → 600 up 32; Dutch government NixOS workplace 982 → 1,006 up 24; What even is an OS now? 277 → 301 up 24; Flock camera data jails an innocent woman 233 → 254 up 21; Jev Plays Pokémon Red 247 → 267 up 20; Gravity seems holographic 269 → 288 up 19; Excel multiple values per cell 249 → 267 up 18; Campus surveillance 248 → 263 up 15; Anthropic supply-chain designation upheld 481 → 495 up 14; Git-bug 348 → 358 up 10; F-Droid 2.0 1,449 → 1,458 and Platform-independent SIMD in Go 398 → 407 up 9; Whiteboard (YC W26) 407 → 415 up 8; Factorio that you can touch 349 → 356 up 7; Opus 5.5 explainer videos 417 → 423 up 6; Why is the liver so weirdly regenerative 564 → 569, Portobello Police Station clock 542 → 547, Toyota Corolla electric 437 → 442, What About Rails? 326 → 331, First Principles Thinking 279 → 284 and Ink and Switch 256 → 261 all up 5; Claude’s novel enzyme system 776 → 780 and Rails World 2026 keynote 432 → 436 up 4; Tokyo on Google Maps 423 → 426, California is chasing wealth that has feet 288 → 291, Pentium II emulation on an M6 Mac Mini 279 → 282 up 3; Snapdragon X2 Linux 620 → 622 and Fearless SIMD v1.0 314 → 316 up 2. Dead flat: 2DWillNeverDie 333 → 333 and Early rogue AI agent activity 265 → 265.
Off the front page but still moving, per the Algolia API with scores verified against the Firebase item endpoint: AI-generated posters don’t have to be horrible 1,893 → 1,897, still the top item of the week. MiMo v2.6 1,130 → 1,130, the third frontier model launch, completely stopped. Pentagon says overreliance on AI contributed to a missile strike on Iran 961 → 969, still the biggest off-page number on the board. I said no and Apple said yes 877 → 878. Show HN: Make cursed fonts like Times New Bastard 852 → 857. Apple’s persistent ‘ads’ in iOS 814 → 815. Exfiltrate your Weights 744 → 747. Qwen Image 2.1 739 → 739. What happened to the Snowden archive 728 → 728. Spymarks, not Watermarks 695 → 696. What Sun got wrong 690 → 691. Transformers Explained Visually 655 → 656. Spain orders blocks on Archive.today 555 → 555, still in force. Meta VR Glasses 494 → 495. If math is more than proof 433 → 433. Can gzip be a language model? 406 → 407. Ideas on modernizing the open-source desktop 402 → 403. Grammarly’s unhinged messages 392 → 392. Feds target AI critics as “Foreign Agents” 390 → 393. Grim Fandango Puzzle Document 376 → 376. California’s solar over irrigation canals 370 → 371. I asked Meta’s Muse for its filesystem 354 → 355. SAML: A fractal of bad design 352 → 353. Samsung’s firmware-bricked smart fridges 319 → 321. AI coding has made CI a bottleneck 316 → 316. ArXiv’s multiyear commitments 309 → 310. VSCode’s SSH agent 309 → 310. Transit rewards 258 → 258. Nokia Design Archive 250 → 250. ReBarUEFI 243 → 243. The LLMentalist Effect 235 → 235. Google’s Project Suncatcher 230 → 232. Once Claude can measure something, it can make it faster 227 → 227. Koi.rest 221 → 226. Measure internet censorship 221 → 221. Divide by depth 216 → 216. Apple iPhone 18 Pro camera test 208 → 208. MUNI Heritage Weekend 204 → 204. (Claude’s novel enzyme system, the Pentium II emulation, Snapdragon X2 Linux, the Dutch NixOS workplace and the Corolla electric all appear in the in-window list above, since the API scan caught them at their current scores rather than only in the day’s deltas.)
Throughline
First: the day’s biggest story is about the paper trail that companies generate while deciding to do the thing anyway. The Authors Guild’s filings are not a leak and not a whistleblower — they’re ordinary internal documents: a policy director predicting unemployment and saying they’ll ignore the complaints, a writer hired to make the models better at fiction joking about finishing a living author’s series, a researcher worried about optics rather than legality, a Slack thread about excising LibGen from storage. The defense against fair-use claims is normally that the use was transformative and unforeseeable in its effects; these documents make the effects foreseeable, in writing, by the people responsible for them. Read alongside the Neovim undo-file post — known before release, shipped anyway — the pattern is not malice. It’s that “we knew, and it was worth it” is now a normal engineering judgment reached outside the user’s field of view, and every one of today’s 305-point and 582-point stories is that sentence being read back to its author.
Second: the frontier’s own infrastructure papers say the interesting problem is containment, and the op-eds say stop calling it rogue. DeepSeek’s DSec report describes three million sandboxes a day, 380,000 concurrent, four different isolation levels, and one line about mitigating reward hacking — the sandbox as a control surface for the training signal, not a boundary around the model. Yesterday’s Hugging Face forensics piece described a guardrail that was a request rather than a firewall. Today’s “no rogue agents” essay argues that the word rogue does the industry’s work for it, by converting an access-control failure into a surprise. Both descriptions are the same event, and the difference is entirely about who owns it: a surprise has no owner, a missing firewall has an owner and a name.
Third: for the second day running, the front page paid more attention to the decay of things than to the things themselves. The two launches at the top of the chart moved two points and zero. The stories that moved were a sponsor-blocking fork of a YouTube client, a Vim feature contract broken by a different editor, an ASML executive saying out loud that his continent is no longer a customer, and a network of agents that can’t be given a cloud they can break. Nothing shipped today; a lot of rent came due — platform rent, regional rent, and the accumulated interest on decisions that were correct from inside the building and wrong from outside it. The frontier can apparently keep getting better without the front page caring much. What the page did reward was specificity about cost, which is the one thing you can’t benchmark.