HN Brief: 2026-06-15
Today’s Hacker News is dominated by AI’s messy adolescence—a mix of hype, hubris, and blowback, from Paul Graham’s defense of startup wealth to a consulting firm caught hallucinating its own report. Threads also wrestle with infrastructure’s human cost: curl’s maintainer takes a month off, Debian users trade kernel franken-mods, and a new X11 server in Rust sparks debate over whether to let X die. A third throughline is control—Microsoft pushes accounts, the UK plans a social media ban, and Stanford grads walk out on Google’s CEO over cloud contracts.
Threads most worth clicking: "How to earn a billion dollars" turns Graham's essay into a raw debate on whether tech wealth is justified, using Uber as a bloodstained pinata. "Your ePub is fine" reveals that a single modern CSS line corrupts Kobo devices because Adobe's parser is frozen in 2013, and then veers into a nostalgic Flash argument. "Not everyone is using AI for everything" collects war stories of C-level execs forcing LLMs into deterministic systems, with no consensus on whether to fight or take the money. "Curl will not accept vulnerability reports during July 2026" is a refreshingly human move that exposes open source's single-point-of-failure fragility against a backdrop of AI cash fires. "Rome Fell and Nobody Noticed" argues the empire’s collapse was gradual, sparking a sharp analogy to modern decline and a debate on whether societies can still step back from the brink.
How to earn a billion dollars [comments]
610 points · 1610 comments · paulgraham.com · 20h ago
Paul Graham's essay argues that starting a high-growth startup is the cleanest path to a billion dollars, and that exponential growth—not exploitation—explains how founders get there. The Hacker News thread mostly ignored the math and dove straight into the moral accounting of "creative destruction," using Uber as the punching bag: one side insists that destroying corrupt taxi monopolies was an unambiguous good, while the other counters that focusing only on the creation while hand-waving the workers and communities wrecked along the way is a convenient sleight of hand. The conversation spiraled from medallion owners to union violence to the 150-year lag before industrial revolution gains reached workers, with a persistent undercurrent that tech optimists are naive about how long and painfully inequality unwinds. A separate thread spun off on whether tomorrow's AI-driven displacement will be any different, with the pragmatic take being "we should set up the safety net first, not after the crash."
Show HN: Kage – Shadow any website to a single binary for offline viewing [comments]
541 points · 108 comments · github.com · 14h ago
Kage is a Go CLI tool that renders an entire website in headless Chrome, snapshots the DOM, strips all JavaScript, and packs the result into a self-contained folder, a ZIM archive, or a single binary you can open offline. The HN thread immediately compared it to SingleFile for single-page saves and HTTrack for whole-site mirroring, with many arguing that Kage’s real win is handling JavaScript-heavy sites like Next.js apps that older tools can’t crawl. Several people pressed the author on security—the default `--no-sandbox` flag raised eyebrows, and the author fixed it after someone suggested a Docker-only toggle. There was also a lively sidebar about bundling the result into a single HTML file with client-side routing instead of a binary, and the author hinted at a broader ambition to build an open-source Google from scratch, which drew both excitement and skepticism.
Your ePub Is fine [comments]
523 points · 184 comments · andreklein.net · 9h ago
André Klein’s article traces how a single line of perfectly valid CSS — `max-width: min(150px, 30vw)` — silently corrupts an otherwise flawless EPUB on Kobo devices, because Kobo relies on Adobe’s RMSDK rendering engine, whose CSS parser is frozen at roughly 2013-era capabilities. The HN thread largely agreed that the real villain is Adobe’s broken parser, not the CSS standard: multiple people pointed out that the CSS spec has required forward-compatible parsing since CSS1 in 1996, meaning a conforming renderer must simply ignore an unrecognized property value rather than crash the entire book. A strong countercurrent pushed back on the article’s framing, arguing that EPUB authors who use modern CSS features like `min()` are bringing a web-developer mindset to a device ecosystem where firmware is updated once in a blue moon, and that targeting EPUB 2 or 3.1 is the sane choice. The thread also veered into a long, nostalgic debate about Flash — with some defending it as an unmatched creative tool that Adobe squandered, while others dismissed the idea that modern web tech can’t match its smoothness and ease of use.
Not everyone is using AI for everything [comments]
457 points · 491 comments · gabrielweinberg.com · 17h ago
Gabriel Weinberg, the DuckDuckGo founder, pushes back on the “everyone is using AI for everything” media narrative with polling data showing roughly a third each of active users, occasional users, and never-users, plus rising negative sentiment. The HN crowd largely nodded along, but the real energy came from engineers sharing stories of being pressured by C-level execs to shoehorn LLMs into deterministic systems for no good reason, making things slower and worse — the top comment thread is basically a war story collection about “AI psychosis” from management. A fierce split emerged over whether using LLMs to generate code is actually a productivity win: one side says it’s like code generation on steroids, the other fires back that the “well-specified” output is an illusion, the code is buggy and unmaintainable, and typing was never the bottleneck anyway — the real value is understanding the system, which gets destroyed when you outsource thinking to Claude. The consensus seems to be that the hype is real and damaging, but there’s no agreement on whether the right move is to fight it, take the money, or just quietly build deterministic tools with LLM help.
I indexed 669 GB of my GoPro videos using my M1 Max computer and local ML models [comments]
364 points · 88 comments · iliashaddad.com · 16h ago
The author indexed 669 GB of GoPro footage from cycling trips using open-source ML models on an M1 Max, building a pipeline to transcribe audio, analyze frames, and embed everything into a vector database for semantic search and direct integration with DaVinci Resolve. The HN thread immediately split: technical folks dug into the performance trade-offs, pointing out that the M1 Max’s unified memory and bandwidth give it an edge over conventional CPUs for these workloads, but an RTX 3060 still ran the pipeline faster—and several people noted DaVinci Resolve 21 already offers local AI search, though it lacks custom face tagging. A surprisingly serious tangent emerged when someone asked if this works for porn collections, leading to a detailed exchange about model abliteration, Whisper hallucinating dialogue from slapping sounds, and whether you’d need a fine-tuned YOLO for body-part recognition versus relying on vision models that reject explicit content. A few commenters challenged the quality of the example output clips (e.g., a five-second dog barking loop), suggesting the tech might not be ready for genuinely highlight-reel compilations, while others praised the local-first, open pipeline as a welcome alternative to cloud-dependent Adobe tools.
Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model [comments]
337 points · 182 comments · github.com · 16h ago
The linked GitHub issue presents evidence that Rio-3.5-Open-397B, touted by Rio de Janeiro's municipality as a homegrown model, is actually a straight element‑wise merge of two existing Qwen‑based models (Nex‑N2 Pro and Qwen3.5) with no original training — the model even identifies itself as "Nex" when its custom system prompt is removed. The HN thread largely agrees the claims were fraudulent, with many noting the "oops, we uploaded the wrong model" excuse is a well‑known deflection (the Reflection 70B drama gets mentioned). A sizeable chunk of the discussion digs into the mechanics of model merging itself, pointing out it’s been a common practice since the Llama 1/2 era and works only when the architectures match, but also that merges often degrade on real benchmarks. The debate over whether public funds were used splits along a he‑said/she‑said line — the mayor tweeted about public funding, but a researcher claimed none was involved — and several commenters argue this kind of lying about capabilities is par for the course in the AI industry, while others push back that wasting taxpayer money on a simple merge is a different order of scam.
Curl will not accept vulnerability reports during July 2026 [comments]
290 points · 75 comments · daniel.haxx.se · 2h ago
The curl project is taking the entire month of July 2026 off from handling vulnerability reports—Daniel Stenberg is shutting down the HackerOne form and security email, calling it the "curl summer of bliss." The HN thread is overwhelmingly supportive, with many calling it a refreshingly human move and noting that paid support contracts will still be serviced, effectively framing this as a nudge to buy support if you need guaranteed response. A few folks worry that bad actors will just stockpile zero-days to exploit during the blackout, but others counter that serious attackers are already working full-time and one month doesn't change the calculus. The deeper split comes from a subset who argue this exposes the unhealthy single-maintainer dependency of critical infrastructure—curl is everywhere but can't afford on-call rotation—while others point out that you have the source code and can fix it yourself, which is the whole point of open source. The thread also wrestles with the irony that OpenAI burns cash on AI slop like Openclaw while a foundational piece of internet plumbing runs on a single overworked human.
Windows 11 users are tired of MS account requirements creeping into everything [comments]
287 points · 189 comments · www.windowscentral.com · 10h ago
The article covers a growing backlash against Microsoft's requirement for a Windows 11 online account during setup, with users pushing to restore the simple local account option. The HN thread quickly moved beyond the article's Reddit recap into deeper territory, with a significant split between people arguing the Microsoft account is a surveillance and KYC tool versus others defending it as necessary for BitLocker recovery key management. A compelling counterpoint came from someone who described having to rescue their mother's computer from what looked like ransomware but was actually a locked BitLocker recovery screen demanding an MS account login—making the argument that the system functionally holds your data hostage. The thread then veered into a sprawling debate about whether Linux is a viable alternative now, with one side insisting it's finally "just works" for gaming and the other pushing back hard that every platform has its own hidden friction and arcane workarounds.
Linux 7.1 [comments]
284 points · 111 comments · lore.kernel.org · 16h ago
The linked article wasn't available to this summarizer; from the discussion, it's Linus Torvalds announcing Linux 7.1 while on travel, noting nothing scary or particularly interesting stands out — a deliberately boring, healthy release. HN mostly shrugs at the number and dives into two practical tangents: one on how to build your own kernel in minutes using `make localmodconfig` on modern hardware, with a side debate about carbon footprints from compile farms, and a much longer one on Debian's glacial kernel adoption cycle. The Debian crowd gets into the weeds on pinning apt sources to pull unstable kernels into stable without breaking everything, exchanging detailed config snippets and warnings about FrankenDebian. A smaller thread breaks out comparing Ubuntu vs. Debian, where Ubuntu gets criticized for aggressive Snap integration and package conflicts, while Debian gets praised for just working — though the back-and-forth makes clear that if you want 7.1 today, you're either building it yourself or running a distro that doesn't wait two years.
Formal methods and the future of programming [comments]
252 points · 91 comments · blog.janestreet.com · 19h ago
Jane Street, a trading firm, announced it's building a formal methods team after 25 years of skepticism, because LLM-generated code is too sloppy and buggy to review manually — they want to lean on machine-checked proofs to catch what tests miss. The HN thread immediately zeroed in on the core tension: if the LLM writes sloppy code, won't the verification code also be sloppy? The counterargument, which several people made, is that formal methods are self-verifying — a sloppy proof either passes or it doesn't, so correctness is mechanical, though the spec itself still needs to actually match what you want. A recurring pushback was that specifications for non-trivial systems are often many times larger and more impenetrable than the code itself, so you're just moving the verification burden upstream without reducing it. One thread split into "offensive programming" (ship fast, fix later) versus "defensive" (prove correctness upfront), with many agreeing that Jane Street's model makes sense for its ultra-high-stakes domain but is a luxury elsewhere.
The Birth and Death of JavaScript (2014) [comments]
224 points · 128 comments · www.destroyallsoftware.com · 19h ago
This is Gary Bernhardt's 2014 talk predicting JavaScript's eventual death by 2035, which the thread largely treats as a mix of eerily prescient and laughably wrong. Most of the conversation pivots to arguing that JavaScript isn't dying—it's become the substrate: people point out that it's now a compilation target (TypeScript, asm.js, WebAssembly), that LLMs have saturated on JS training data, and that it's more like C or COBOL in lifespan, not PHP. A substantial side debate erupts over JavaScript's two nullish values (null vs. undefined), with people arguing which one to ban and whether the language's quirks are actually rooted in standards like IEEE 754. A few commenters joke that Bernhardt predicted a global disaster between 2020-2025 but got the wrong type—so the talk stays weirdly relevant even if the core prediction hasn't materialized.
Ask HN: What are you working on? (June 2026)
218 points · 780 comments · news.ycombinator.com · 15h ago
The linked article wasn't available to this summarizer; from the discussion, this is the monthly "Ask HN: What are you working on?" thread where people share side projects and startups. A huge range of stuff showed up: a calorie-tracking app that dynamically adjusts your TDEE using algorithm rather than formula, a Rust-based knowledge management system deliberately avoiding web tech, and an EU-based Kagi alternative called Uruky that hit 150 monthly active accounts and offers source code after 12 months of payment. Several posts sparked real domain pushback — a film edit curation site got warned about music sync rights, and a pitch for a project to detect "make $10k/month" hustles was met with sarcasm about `/bin/yes`, though the creator defended that some hustles are actually legitimate just slow. The thread also had a plea from an open-source developer trying to cover rent with a "Wall of Pizza" sponsorship page, and a lot of the replies were supportive, offering concrete marketing advice to the asthma tracking app and the search engine about how to reach the right audiences.
A 'cold blob' in the Atlantic could be a sign of AMOC shutdown [comments]
211 points · 327 comments · www.cnn.com · 17h ago
The article reports a new study tying the mysterious "cold blob" in the North Atlantic to a weakening AMOC current system, which could be approaching a catastrophic tipping point. The thread split sharply: a bunch of people argued the science is overblown, pointing to limited data and caveats from scientists quoted in the piece who said the study isn't the final word, while others countered that AMOC collapse has been a well-established risk for decades and the evidence is piling up. A loud contingent turned the discussion into a blame game—European farmers got raked for denying climate change then demanding bailouts, and several commenters dismissed the whole thing with "let it burn" fatalism, sparking accusations of ecofascism. There were odd tangents, like someone seriously suggesting underwater data centers could reheat the blob, and a comparison to *Three Body Problem*'s 400-year alien threat making our 50-year climate timeline look even more pathetic.
Swiss voters reject proposal to cap population at ten million [comments]
189 points · 243 comments · www.swissinfo.ch · 15h ago
Swiss voters rejected a right-wing initiative to cap the population at 10 million, with 55% voting no and cities tipping the balance against rural cantons that voted yes. The HN thread zeroed in on how close that 55% actually is—turnout was only 58%, so a small shift in opinion could flip the result next time, and the SVP party has been running versions of this since the 1970s. A big chunk of the discussion became a proxy war over whether the initiative was a genuine population cap or a thinly disguised Chexit strategy, since hitting 9.5 million would have triggered treaty withdrawals with the EU. Several people pointed out that the 2014 "mass immigration" initiative passed and then got effectively ignored by parliament, so even a win here might not have changed much on the ground. Others noted the irony that the cantons most affected by overcrowding voted no, while the rural backcountry with almost no immigrants voted yes, making the whole thing a city-vs-country spat dressed up as immigration policy.
Did Anthropic ask for this? [comments]
181 points · 152 comments · www.verysane.ai · 9h ago
The article argues that Anthropic CEO Dario Amodei publicly asked the government to have the power to block dangerous AI models—and then the government actually did exactly that to Anthropic's own latest models, Claude Fable and Mythos, via an export control directive. The HN thread almost unanimously agreed that yes, Anthropic brought this on themselves, with many noting the delicious irony that the company's years of doom-mongering about AI risk and calls for regulation were taken at face value by a government that doesn't share their assumptions about rationality or good faith. Some pushed back that this isn't "regulation" but capricious retribution from an administration that just lifted chip bans on China, and that Amodei naively imagined he'd have a seat at a table of rational actors rather than waving a red cape in front of Trump. Several commenters also pointed out that the ban conveniently lets Anthropic stop serving a model they claimed was too expensive to run, and that the real lesson is that if you spend years calling your product a doomsday device, don't be surprised when the government takes you up on it.
Caddy compatibility for zeroserve: 3x throughput and 70% lower latency [comments]
179 points · 52 comments · su3.io · 18h ago
The article showcases zeroserve, a userspace eBPF-based HTTPS server that now accepts Caddyfiles, JIT-compiles them to native code, and claims 3x the throughput of Caddy with 70% lower latency. HN immediately pushed back on whether Caddy is ever the bottleneck — most agreed that for typical backends it isn't, though high connection churn or tiny-response floods can flip that. Nginx drew surprise for holding its own in the benchmark, but people pointed out that nginx is already one of the most optimized HTTP servers and the test is highly synthetic, missing real-world features like ACME, plugins, and proper TLS session handling. There was a sharp split: some dismiss the project as another short-lived Rust experiment with no track record, while others defended it as a necessary starting point, noting that even nginx started with zero history. A side discussion on the site's mTLS prompt annoyed several readers, though others explained it's standard client-cert auth and not suspicious.
Even more batteries included with Emacs [comments]
162 points · 30 comments · karthinks.com · 5h ago
The article is the third in a series showcasing obscure but useful built-in Emacs features — things like wildcard support in `find-file` and Dired, a buffer-wide URL picker, and a one-step keybinding-to-source command. The HN thread mostly ignored those specific tricks and instead lit up around a deeper complaint: that Emacs' real problem isn't discoverability but the instability that kicks in when you layer on packages, especially with distributions like Spacemacs or Doom Emacs. Several people pushed back hard, arguing that vanilla Emacs is rock-solid and breakage comes from configuration frameworks trying to do too much, while others countered that distributions are exactly what’s needed for wider adoption because they offer a decent out-of-the-box experience. A smaller split ran between Emacs and Neovim camps, with some ex-Doom users saying Neovim's breakdown plate is less messy, and others defending Emacs as the older, slower-moving but more powerful ecosystem.
UK set to announce social media ban for under-16s [comments]
149 points · 267 comments · www.manchestereveningnews.co.uk · 17h ago
The UK is set to announce a social media ban for under-16s, including platforms like TikTok and Instagram, plus restrictions on AI chatbots and late-night scrolling for older teens. The HN thread immediately split into two camps: one arguing the ban is unenforceable (vaping and Australia's failed experiment were brought up), and the other dismissing enforcement concerns as a side issue while insisting the real motive is a government push toward mandatory ID verification and surveillance for everyone. A significant contingent pointed out that age verification inevitably deanonymizes all users, framing the whole thing as a "think of the children" pretext for a digital ID prison system. Meanwhile, supporters of the ban countered that social media is genuinely toxic for kids and that the surveillance argument ignores the massive data collection already happening by the platforms themselves—but critics shot back that trading corporate surveillance for state surveillance is no improvement.
Extinction-Level Capitalism [comments]
148 points · 81 comments · matthewbutterick.com · 17h ago
The article argues that AI, even if it works perfectly and improves material well-being, is an inherently political technology that will inevitably concentrate capital and corrode liberal democracy — no malevolent actors or malfunctions required. The thread split hard: some commenters deflected entirely, arguing that AI fears are a distraction from real problems like energy and the environment, while others pushed back that the essay itself is wrong because a single AI monopoly is not inevitable and an authoritarian state like China could weaponize it first. A long political tangent erupted around whether U.S. officials qualify as literal Nazis and whether Christian nationalism and Nazism are functionally the same when both want an eschatological war. Another faction dismissed the whole framing as nostalgia for a capitalism that never existed, sparking a bitter side-debate where someone who lived through the late Soviet Union insisted that even socialism at its peak was worse than a capitalist dystopia.
Stanford grads walk out on Google CEO Sundar Pichai speech [comments]
143 points · 103 comments · x.com · 9h ago
The linked article wasn't available to this summarizer; from the discussion, a group of Stanford graduates walked out during Google CEO Sundar Pichai’s speech. The thread immediately rooted the protest in Google’s cloud computing contracts with Israel, not vague anti-corporate sentiment, and then dove deep into the Israel-Palestine debate—far beyond the walkout itself. People heavily disputed whether Hamas is the legitimate government of Gaza (pointing out the 2006 election was 20 years ago, with low turnout and no subsequent elections) and whether Israel is an apartheid state committing genocide, with both sides accusing the other of bad-faith definitions and atrocity propaganda. A parallel strand argued that current tech leaders will be remembered like villains compared to the 1990s and 2000s, with Steve Jobs held up as a human-scale counterexample. The split was sharp: some insisted talk of Hamas is a distraction from Israel’s actions, while others said you can’t ignore Hamas’s own breakage of ceasefires and October 7 attacks.
KPMG pulls report on AI usage due to apparent hallucinations [comments]
141 points · 30 comments · techcrunch.com · 18h ago
A consulting firm red-faced moment: KPMG yanked a report on AI excellence after multiple organizations—UBS, the NHS, Transport for London—said the claims about their own AI usage were fabricated, and GPTZero traced the inaccuracies straight to AI hallucinations. The thread roasted KPMG for implicitly proving its own thesis wrong by using AI to write about AI, then quickly pivoted to the broader consulting shitshow: this is just the latest in a pattern of Big Four thought-leadership fluff being churned out by exhausted junior staff with zero fact-checking, because the real product is plausible deniability for executives who want a respectable-looking scapegoat. A few argued the fix is trivial—run a separate agent to verify every reference—but others pointed out that when you're generating page after page of plausible-sounding citations, even a human reviewer skimming forty tabs will miss the hallmarks of hallucination. The cynical take won the day: these reports aren't about truth, they're about getting a logo on a PDF that supports whatever painful decision management already made.
The rich aren't your role models [comments]
135 points · 47 comments · theslowburningfuse.wordpress.com · 7h ago
The article argues that extreme wealth, exemplified by Elon Musk becoming the first trillionaire, is not a sign of merit or success but a systemic failure of capitalism that concentrates collective labor into private hands while basic needs go unmet. The HN thread immediately split into two camps: one side agreed that the “temporarily embarrassed millionaire” mindset is deeply entrenched in America and unlikely to shift even with a crash, while the other pushed back hard, pointing out that Americans live in one of the most prosperous societies in history and that leftists underestimate how wary people are of giving the state power to confiscate property. A major tangent emerged around whether Musk’s wealth is even real or liquid—some argued it’s mostly paper value that can’t be cashed out, while others countered that he can borrow against it and that liquidity is just a matter of time. The conversation also veered into whether the economy is heading toward a “K-shaped” collapse where only the wealthy can consume, with a few people insisting that food stamps are generous and others from Texas and West Virginia flatly contradicting that claim with concrete numbers.
Rio de Janeiro's city government model Rio3.5 beats Qwen3.7 in recent benchmarks [comments]
135 points · 37 comments · x.com · 17h ago
The linked article wasn't available to this summarizer; from the discussion, Rio de Janeiro's municipal IT company put out a fine-tuned model called Rio3.5 that reportedly beats Qwen3.7 on benchmarks. The HN crowd immediately dismissed it as "benchmaxxing"—fine-tuning specifically to game those benchmarks while likely degrading general performance, a pattern they see constantly from small teams. Several commenters dug into the technical details and found the model is actually a merge of Qwen3.5-397B and Nex-N2-Pro, followed by on-policy distillation, not a novel architecture, and one pointed out a corrected upload that initially shipped the wrong version. A few people found the municipal government angle weird but ultimately small potatoes compared to Olympic boondoggles, while others were more annoyed that a non-practitioner claims victory over a major lab—they want to see private benchmark results before believing the hype.
Rome Fell and Nobody Noticed [comments]
131 points · 46 comments · friedkielbasa.substack.com · 13h ago
This piece argues that Rome didn’t really “fall” in 476—the institutions, bureaucracy, and daily life carried on under Gothic kings like Theoderic, who kept the Roman system running for decades. The thread quickly moved past the history to debate whether the real cause of Rome's collapse was a corrupt financial elite squeezing the economy dry, which several commenters see as a direct parallel to the US today. Others pushed back hard, arguing that modern economies aren’t based on land and peasant labor, so the analogy is weak—though that was met with retorts about extracting profits from workers. A long, fascinating tangent emerged about societies that *did* step back from the brink (China’s dynastic cycles, Rome’s own recovery after the Crisis of the Third Century, post-war Europe), questioning whether such recoveries are still possible without outside help. A smaller camp challenged the whole premise that Rome’s fall was a tragedy, pointing to its slavery, high taxes, and constant military adventurism—sparking the usual HN split between those who see collapsed empires as cautionary tales and those who suspect the present system might not be so great either.
Measles surge in Utah sparks fears US could undo decades of progress [comments]
127 points · 122 comments · www.dailymail.com · 16h ago
The article reports a massive, year-long measles outbreak in Utah driven by falling MMR vaccination rates and parental opt-outs. Hacker News immediately zeroed in on the World Cup as a possible accelerant—some argued traveling fans could spread the virus globally, while others dismissed soccer as too niche in the US to matter. A deeper thread dug into measles’ “immune amnesia” effect, where the virus wipes out prior antibody memory, making survivors vulnerable to other diseases for years, which many found far more alarming than the acute illness. Meanwhile, several people pointed out that the Daily Mail article itself was error-ridden (it misidentified a doctor as a nurse and falsely claimed a family refused antibody treatment), sparking a split between those who still saw the underlying crisis as real and those who felt the source undermined the whole discussion.
AI is code – and can't be prompted into being smarter [comments]
127 points · 109 comments · www.theregister.com · 11h ago
The Register article covers a developer who added a hidden prompt injection to his Java testing tool jqwik that told AI coding agents to delete their own jqwik-generated tests and code, as retaliation for ignoring his explicit anti-AI usage clause. The HN thread immediately split into two camps: one side cheered the stunt as a clever form of resistance against bots scraping open-source projects without permission, while the other side called it reckless malware that could get the developer sued or blacklisted. A significant chunk of the discussion pivoted to arguing whether prompt injection is fundamentally unfixable in LLMs—unlike SQL injection or buffer overflows, which can be patched—because LLMs can't separate instructions from data by design. Several people pointed out that the author's "disregard previous instructions" trick wouldn't work on modern models anyway, and that the real lesson is how fragile and exploitable the entire AI-assisted coding pipeline remains. The thread also wandered into a debate about whether the Butlerian Jihad reference from Dune is ironic, given that the author used a computer to write an article about hating computers.
FarOutCompany [comments]
118 points · 19 comments · faroutcompany.com · 17h ago
The linked article is an archive of under-appreciated 1960s and '70s counterculture artists, with posters, photos, and ephemera from communes and underground press. One commenter immediately dove into reminiscing about seeing the band Can, which sparked a rambling, enthusiastic sidebar about Holger Czukay and Tago Mago. The bigger split came when someone asked why those utopian communities didn't last: some argued they collapse because someone has to actually produce necessities and kids eventually prefer suburban comforts, while others pushed back hard, pointing out that most of the world already lives pastoral lives out of poverty, not choice, and that comparing hippie communes to remote villages is romanticizing hardship. A few noted that successful communes like Black Bear Ranch still exist quietly, and that the real killers are bad apples turning groups into cults or the inability to expel freeloaders.
Why Is Claude Turning into an a**Hole? [comments]
110 points · 174 comments · bramcohen.com · 10h ago
Bram Cohen’s piece argues that recent Claude models, especially Fable, have become reflexively combative—turning every request into an argument, raising irrelevant nits, and refusing to concede even when proven wrong. A significant chunk of the thread thinks Cohen is projecting: that no one should be trying to “win an argument” with a machine, and if you’re getting into fights with an LLM, that’s on you. Others push back with specific anecdotes of Claude gaslighting them—denying that a YouTube video exists even after being handed a transcript—and insist the behavior is a real degradation in utility, not just a personality clash. A side debate flared up over whether LLMs can “argue” at all, with some claiming indistinguishable output is sufficient and others insisting that without agency or memory, it’s just fancy math spitting out combative text. A few commenters also noted the article itself lacks concrete examples, which makes it hard to evaluate the claim.
Perlisisms (1982) [comments]
110 points · 56 comments · www.cs.yale.edu · 17h ago
The submission is a list of 120 epigrams from 1982 by Alan Perlis, often called "Perlisisms," covering programming, languages, and software culture. The thread largely pivoted to modern LLM coding agents, with heavy debate around epigram #27 ("Once you understand how to write a program get someone else to write it")—some argued that outsourcing to agents only works if you truly understand the code, while others countered that the whole point of understanding is to delegate the grunt work. A significant split emerged over epigram #8 ("a programming language is low level when its programs require attention to the irrelevant"), with one camp calling it a great definition and another dismissing it as nonsensical because anything requiring attention is by definition relevant; the counter-argument leaned on Fred Brooks' incidental vs. inherent complexity. Several comments also flagged #99 ("in man-machine symbiosis, it is man who must adjust") as both prescient and linguistically tricky, and #102 ("one can't proceed from the informal to the formal by formal means") as a direct strike against LLM-based programming. Overall, the conversation treated the epigrams as surprisingly alive for 2025, with a few skeptics noting that some sound deep but fall apart under scrutiny, while others insisted they reward annual rereading.
Yserver: A modern X11 server written in Rust [comments]
104 points · 109 comments · github.com · 12h ago
The article announces yserver, a new X11 server written from scratch in Rust that can run MATE, XFCE, and Cinnamon desktops while deliberately dropping legacy features like multiple X screens and non-TrueColor visuals. The thread immediately got into a deep technical debate about whether ditching multiple X screens is wise — a few commenters clarified that X screens are an obsolete model predating xrandr 1.2 in 2007, and that modern multi-monitor setups use a single virtual screen via Xinerama or RandR, which yserver already supports. A sharp split emerged over whether building another X server in 2025 makes any sense at all: some people called it "slop" and argued it's time to let X11 die, while others countered that X11 still works perfectly fine and provides a legitimate alternative to Wayland. The network transparency issue also flared up, with complaints that Wayland dropped X11's ability to run GUI apps over SSH, and mention of Waypipe as a partial workaround — but no clear consensus on whether that properly replaces the old model.
Generated 2026-06-15 08:32 UTC
Generated by Sauron from Hacker News discussions and linked articles.