Top of Hacker News

Last week's climbers: iroh, a Rust stack that dials devices by cryptographic key instead of IP address (and just shipped 1.0); Plane, a self-hostable, air-gap-capable alternative to Jira and Linear; Presidio, Microsoft's PII scrubber that teams now wire in front of every LLM call; Meshery, a manager for your own cloud-native infrastructure.

The odd one out is freeCodeCamp, at 450k stars and trending yet again with nothing new to say. iroh reached 1.0 the hard way: it tore out its QUIC engine and shipped one it wrote itself.

In today's GitHub Trending: activepieces, the open-source Zapier alternative that now exposes your automations to Claude and Cursor as MCP tools; TinaCMS, a Git-backed CMS you edit visually on the page; System Informer, a Windows process monitor; and Media Downloader, one GUI over yt-dlp and a dozen CLIs.

The odd one out is international_results, a CSV of every international football match since 1872, with zero Hacker News or Bluesky discussion. It's trending because the 2026 World Cup is in its group stage, not because anyone recommended it.

Capcom построила Requiem вокруг переключения между двумя героинями: за Грейс это тихий, нервный стелс, за её напарником — открытое действие. Один и тот же кошмар вы проживаете двумя разными способами.

Часть игроков предпочла бы две отдельные кампании вместо постоянных переключений. Но даже они сходятся в одном: как хоррор игра бьёт точно в цель.

Парадокс в том, что даже авторы гневных отзывов оговариваются: «игра хорошая». Спорят они не о механиках, а о тематике события.

А те, кто пишет про сам геймплей, советуют его едва ли не хором: трёхполосный бой в ритм оказался на удивление затягивающим.

Ответ сообщества почти единодушен: контента и новизны маловато, карты тесные, а полная цена не оправдана.

Но тот же хор добавляет: на хорошей скидке отстреливать врагов фирменным рентген-выстрелом по-прежнему весело.

Новички в отзывах описывают один и тот же перелом: на высоком ранге игра перестаёт прощать ошибки и впервые становится по-настоящему страшной.

Ветераны добавляют отрезвляющее: гринд за снаряжением не заканчивается никогда. Для кого-то в этом и есть весь смысл.

Совет, который повторяют ветераны: не смотрите гайды и видео. Без карты и подсказок первый час по-настоящему тревожен, и именно в этом вся соль.

Многие признаются в страхе перед открытой водой и говорят, что игра дала пережить его, не заходя в реальный океан.

Photocopies, wallets and work badges all get rejected; only a passport, license or national ID plus a live selfie clear the gate. The stated goals are to prevent abuse, enforce policy and meet legal obligations, and the check can fire mid-session.

Anthropic never holds the images: Persona keeps them encrypted and walled off from training. What looked like a fresh crackdown turned out to be an old page resurfaced, and the suspicion was less about privacy than who is leaning on Anthropic to build the gate.

Cover letters, résumés, scripted "tell me about a time" answers, even live video calls: each is now, in the authors' words, "infinitely scalable and practically free." Strip the signal from both sides and you select for talent at being hired, not at the job.

The fix on offer is friction: sudden constraints, a defended tradeoff, a culture prizing honesty over polish. The thread shrugged. Qualified people were filtered out by invisible wires long before any model arrived, and the cleanest counter-signal turned out older than software.

The trap is a sunk-cost spiral. Each near-miss requirement gets bolted onto the abstraction rather than rejected, until the code is a tangle of booleans nobody can reason about. Her slogan: "the fastest way forward is back." Inline it into every caller, prune dead branches, rebuild.

The thread split on whether this survives real teams. One camp says it gets weaponized to bless permanent slop; another calls the framing a category error. And when a model maintains the copies, sameness that drifts silently becomes its own hazard.

Apertus ships in 8B and 70B sizes, claims more than 1,000 languages, and bakes in EU-AI-Act compliance: PII removal, opt-out respect, memorization prevention. The pitch is total reproducibility — "to AI as Open is to Source" — not benchmarks. Cutoff: March 2024.

One founder reports building real RAG software on it today; it is not ready for agentic work, but it drives. Skeptics aimed at the institution: a model moving at committee speed, a country with no domestic GPUs, and a project whose best output may be the people it trained.

A single 2014 internal email reframed everything: "Right now we need 2 more companies to cover our costs." If a startup is kept alive only to justify the fees charged against it, was its work ever real?

The author's verdict cuts the other way. The core database tech was real and his code shipped; the chef-and-cleaner expenses were someone else's scheme. The fraud sat above him, not in his commits.

The thread surfaced uglier cases: engineers who had unknowingly stood inside far larger frauds, from markups to a $10B telecom collapse.

Price a $200k engineer at roughly $96 an hour and the math gets blunt. A $400/month tool needs only about four maintenance hours a month to justify rebuilding; a $25k/month Salesforce contract implies about 1.5 engineers to stand still.

So the cheap tools fail build-vs-buy and the expensive ones pass it, the opposite of intuition. The post doubles as a launch for River, Brandur's open-source queue.

The thread's counter: the real cost was never the features. It's SSO, audit logs, and the integration web — the moat an LLM won't clone.

Hotz contrasts Anthropic's policy-and-danger output with rivals' plain engineering posts, and reads recursive-self-improvement warnings as marketing copy. The doom, he claims, values the industry on hypothetical future capability, not today's.

The thread's best objection turns his evidence on him: if institutional investors think mass unemployment is nonsense and invest accordingly, where is the doom-to-valuation link?

Another notes the export-control fight makes danger talk expensive, not profitable; one calls the post projection.

Adoption crept up almost linearly, roughly 10 points every three years since 2018, which reads like failure until you notice it's the curve of a mature protocol driven by ISPs making rational cost decisions. Jio, Vietnam and Saudi Arabia climb far steeper.

China's traffic goes unmeasured, so the true global figure is likely higher, perhaps 65 to 75 percent. The long timeline isn't a stall; it's winning slowly.

The thread torched the lazy explanations: no, the growth isn't just AWS bots, and no, you couldn't have "just" extended IPv4.

Roughly 99.9% of ATProto accounts sit on Bluesky's own servers and never register a backup rotation key. That single gap turns the operator into a custodian who can impersonate you across Tangled, Grain, and Leaflet, not just one app.

The theory said you would hold the private keys. The product decided otherwise, because making millions of people manage cryptographic material would have crushed signups.

The fix is unglamorous: make backup-key enrollment mandatory at account creation, baked into the client.

PRINCE is a Bayer preclinical-research assistant that grew from keyword search into a conversational agent. Its reliability rests on context engineering (each agent sees only what it needs) and harness engineering (saved state, retries, fallback providers).

The article names RAGAS and faithfulness metrics but reports no accuracy or latency figures of its own.

A reader supplied the missing one from the source paper: an overall capability of 3.1/5.0 — for a medical research tool.

Among US adults, 17% show severe Problematic News Consumption — and 61% of them report feeling unwell, versus 6% of everyone else. Globally, 40% now avoid news, a record high; in Canada, 69% sometimes do.

Negativity bias is the brain weighting and remembering threats more heavily. It evolved when threats were local and actionable. The feed makes both untrue.

The prescription is not abstinence but management: time limits, depth over volume, and linking awareness to action.

Maritime work is measured in ton-miles, not tonnage. Fossil fuels move in huge flows over long distances, so they punch above their weight in freight energy. As coal, oil, and gas demand falls, the bulk carriers and tankers hauling them disappear — a double-dip.

The prescription: subtract that shrinking denominator first, then electrify the rest and reserve biofuels for the residue.

The thread's verdict on the writing was less kind than on the logic.

The vaccine in question is not COVID. It is a seasonal flu shot built on mRNA, which can be reformulated late enough to chase the strains that actually circulate rather than a months-old guess.

The 'agency drama' in the headline traces to Vinay Prasad, an FDA official critics call vaccine-skeptical. The expert committee recommended approval over that internal resistance.

Then the thread turned on the official's defenders, and on the article itself for citing no data.

An analysis of ACA in-network claims puts denial rates between 13% and 35% depending on the insurer. The breakdown beneath that range is the surprise.

Only about 1 in 20 denials cited medical necessity. The rest were administrative: excluded services, missing referrals or prior auth, or a reason the insurer simply never gave.

The thread's twist: 'never specified' is not the same as harmless.

Windows 95 nailed it: a clean dialog that let you pick an installed program inline. That same flow survived through 98, ME, and 2000.

Windows XP added an online lookup that mailed your file extension to a Microsoft web service, now defunct. Windows 10 and 11 flattened everything into low-contrast boxes that surface a Store search first.

The thread fills in the gap the author skipped, and finds a working path he missed.

The single em-dash tell is dead, the author argues. What gives AI away is aggregate sameness across many outputs, not any one artifact.

His proof: a collage of roughly 150 AI-generated kids' encyclopedia covers, many of them category bestsellers, recycling the same tropes and surnames.

The thread named the mechanism, then questioned whether the sample proves anything.

The reframe is tiny: never request approval for work already in your lane. State the action, attach a deadline, and the silence becomes a yes.

It transfers the cognitive load. Saying no takes one click; saying yes makes the manager co-own a decision they would rather not study.

But the same sentence reads as initiative or as a hostage note, depending entirely on whether the thing you announced was a prototype or a push to prod.

John-David Dalton's step-back wasn't a workload story. It was a death, a divorce around 2019, and a long climb back through therapy, exercise, and hobbies that weren't code.

OpenJS's thesis: long-term sustainability beats constant output, and critical infrastructure should rest on committees, not one exhausted human.

The thread's sharpest objection wasn't with that. It was that the foundation flattened a raw phone call into a tidy corporate blog post.

His effort argument is a picture: AI-writing is hooking a motor to a stationary bike and calling it exercise, or taking a helicopter to the summit of Everest.

His trust argument is harsher. "AI helped" is unverifiable, so it's a lie about how much you did. Admit it once and all your future work reads as spam.

The thread's cleanest counter pointed out that he asks you to take his own no-AI claim purely on his word.

The dramatic claim has three sets of hands on it before it reaches the reader. A general told a senator, the senator said it aloud on June 11th, and The Economist printed it. Two of the three relays are a politician and a flag officer.

The briefing's actual headline is about something else entirely: whether blocking Anthropic was arbitrary. The hacking line is the evidence, not the thesis.

And if one model can do it in hours, commenters note, the framing collapses the moment you ask which model couldn't.

It is genuinely free and open source, seven years of community work on a forked engine, with thousands of fully simulated units smashing into deformable terrain.

But a June 2026 publishing deal with Hooded Horse sits under the marketing gloss, and longtime players surfaced both the governance drama and a community that can vote to kick you and seize your units.

The same lobbies, others insist, are where the real charm lives.

Alex Kritchevsky splits the movement into two claims and accepts one. Exterior algebra — wedge products, multivectors, duality — should be taught early. Fine.

The Geometric Product, which fuses a scalar and a bivector into a mixed-grade object with no general geometric meaning, he calls operator composition wearing a costume, useful mainly for rotations.

A reviewer of GA papers and an animation engineer both said: that's exactly the boundary.