HN Brief: 2026-06-18
Today's front page was a tug-of-war between two entrenched camps: the "AI is a turnoff" crowd, who got fresh survey data showing most Americans actively hate the label, and the "AI is a new kind of engineering" crowd, who argued that tools like GLM-5.2 are getting so good at overthinking they've become a liability for real work. A third thread, on the collapse of U.S. science funding, sat awkwardly between them, offering a reminder that the biggest story might be the one where people are simply ready to leave—not just a platform, but the whole country.
"Want your images back? That'll be $5" is a fast, infuriating read on how Photobucket turned old habits into a hostage ransom, and why a chargeback is the only real fix. "MicroUI" is worth clicking for the rare thread where a developer defends a 1,100-line debug library against accusations it's "unusable" for not supporting screen readers—a perfect HN meta-slapfight. "The Competitive Moat That AI Can't Replicate" is worth it just for the thread's verdict that the post itself is AI slop, which makes the whole "human connection" thesis feel less like a deep insight and more like a sincere joke. "A robot is sprinting towards you" delivers the weirdest hook of the day: a battle royale of eleven LLMs, where the best cheap model beat the most expensive by 27x per win.
Lore – Open source version control system designed for scalability [comments]
1109 points · 595 comments · lore.org · 17h ago
The linked article wasn't available to this summarizer; from the discussion, Epic Games has released Lore, an open-source, centrally-managed version control system built to handle the massive binary files and team workflows that make Git struggle in game development. The HN crowd immediately clocked that this is a direct competitor to Perforce, not Git, and the centralization is a deliberate non-goal—meaning there's no argument for it replacing Git in pure software shops, just a better fit for studios shipping multi-terabyte repos of art assets. A solid chunk of the thread pivoted hard into griping about Epic's storefront practices, with people calling the "free games" model a loss leader tied to spyware, though others pushed back that the competition with Steam is healthy. The core technical takeaway that got traction is that Lore uses content-addressed Merkle trees and chunked storage to do what Git LFS does poorly, and the CLI is deliberately less cryptic than Git's, though a few people pointed out that the real test for adoption is a good GUI for artists who don't want to touch a terminal.
Sixty percent of US consumers say 'AI' in brand messaging is a turnoff [comments]
1039 points · 543 comments · wpvip.com · 19h ago
A WordPress-commissioned survey found that 60% of US consumers actively dislike "AI" in brand messaging — the term itself is now a turnoff, not a draw. The thread quickly zeroed in on the semantic rot: "AI" has become synonymous with "LLM chatbot" to most people, which is a feature they didn't ask for and a signal that the product is seeking venture capital vibes rather than solving a real problem. A vocal chunk of the thread argued the real damage is deeper — consumers connect "AI" with plagiarism, environmental destruction, and job replacement, making it a culturally radioactive label that's actively repelling customers. Others pointed out the irony: people only notice badly-done AI, so slapping the word on anything reminds them of every clunky, generic experience they hate, not the invisible background systems that actually work. The article's thesis — that the winning brand will feel "human in the AI era" — got serious pushback from people who said that's marketing nonsense and the real answer is just to shut up about the buzzword and describe what the product actually does for the customer.
GLM-5.2 is the new leading open weights model on Artificial Analysis [comments]
842 points · 408 comments · artificialanalysis.ai · 22h ago
The post is about GLM-5.2, a new open-weights model from Z.ai that now leads the Artificial Analysis Intelligence Index with a score of 51, beating out the likes of DeepSeek V4 and Kimi K2.6 on both raw intelligence and agentic benchmarks, while being significantly more expensive per task at $0.46. The HN crowd immediately pushed back on the benchmark methodology, noting that the reported results use the "Max" thinking setting which burns through an absurd 43,000 output tokens per task—roughly three times what GPT-5.5 high uses—and that the real cost and speed comparison flips dramatically when you drop to the "High" or "Low" effort modes that users actually reach for. A major thread emerged around the toxic culture of overthinking in these models, with people describing an infuriating pattern where the model spends 15 minutes in a "wait, but what if" loop before writing a single file, and the clear diagnosis is that these systems are being optimized for one-shot benchmark performance rather than real multi-turn collaborative work. Some people argued this is actually a feature, not a bug, because both the Chinese and American frontier labs have incentives to hide their real reasoning and burn tokens for profit, while others pointed to techniques like token-banned lists or XML formatting that can tame the waffling.
U.S. science is in chaos [comments]
803 points · 960 comments · www.scientificamerican.com · 22h ago
The Scientific American piece argues that the post-WWII compact between U.S. science and government is broken, using the death of a decade-long NASA satellite project as a case study in how budget cuts and political targeting have starved research programs. The thread split into two camps: one group focused on the human cost, noting the project was "starved to death" by deliberate budget cuts and a shutdown that gave the team no leeway to recover; the other took a harder line, saying scientists are just another set of government contractors whose customers—the electorate and the administration—can change their minds for "retarded" reasons. A recurring debate on the thread pushed back against the idea that this is about saving money, with many arguing it's about suppressing research on politically inconvenient topics like climate change, while others insisted science has always been subject to political whims and that scientists are not a "special caste." One deeper thread turned into a long, messy argument about whether the U.S. response to climate change or the Catholic Church sex abuse scandal sets a worse precedent for how institutions handle systemic problems, drawing in comparisons to Australia and the global impact of methane release. Overall, the consensus seemed to be that regardless of the cause, the effect is clear: the U.S. is now a place where a huge portion of its scientific talent is considering leaving, and the trust between researchers and the government is described as "completely broken."
Volkswagen started blocking GrapheneOS users [comments]
648 points · 389 comments · discuss.grapheneos.org · 16h ago
The linked article wasn't available to this summarizer; from the discussion, Volkswagen's app started blocking users running GrapheneOS (a privacy-focused Android build) using Google's Play Integrity attestation. The HN thread immediately split: a vocal group argued that app developers have no obligation to support niche OSs, and that users who choose GrapheneOS accept broken apps. But the counterargument, which carried the day, was that GrapheneOS didn't need any special support—it has near-perfect Android app compatibility, and VW's move was a deliberate, anti-competitive block using Google's certification gatekeeping, not a technical limitation. A significant tangent pushed for forcing automakers to open up their server-side APIs so third-party clients could fill the gap, noting that VW actually had an API that was recently made far harder to use. Deeper down, the thread devolved into a familiar HN argument about whether allowing custom car operating systems is a safety nightmare or just a smokescreen for manufacturers locking down hardware they don't own.
Want your images back? That'll be $5 [comments]
626 points · 260 comments · www.lutr.dev · 18h ago
The article is a personal rant about Photobucket, which now traps old users into paying a $5 monthly subscription to access their supposedly "protected" images—only for the author to discover his account was actually empty after coughing up the cash. The thread largely pivoted to GDPR workarounds, with several people pointing out that EU data access laws (or even just claiming to be from the EU or a state like Minnesota) would force Photobucket to hand over any actual photos for free, making the whole $5/month hostage situation legally moot. Others pushed hard on chargebacks as the real remedy—fraud is fraud, and a $5 amount is so small the bank almost always sides with the consumer, which hits Photobucket harder than a refund ever would. A split emerged between people who saw this as a funny cautionary tale and those who cynically noted the model works well enough to tempt future projects, though the rejoinder there was blunt: "That’s not providing value, that’s extortion."
Midjourney Medical [comments]
597 points · 421 comments · www.midjourney.com · 6h ago
The linked article wasn't available to this summarizer; from the discussion, Midjourney appears to have announced a project aiming to deploy 50,000 full-body medical scanners globally to generate a billion scans a month for AI training. The dominant reaction on HN was that this was a joke or a parody of tech hype, with people immediately tearing into the scale—pointing out that a billion monthly scans would require scanning the entire US population several times over, which makes the proposal more about selling luxury gadgets to the ultra-rich. A fierce debate broke out over whether extremely frequent full-body scans would actually be useful for health or harmful, with one camp arguing early detection saves lives while a counter-camp, drawing on medicine and statistics, insisted that massive false positive rates lead to unnecessary surgeries, ruined mental health, and that "more data" doesn't automatically mean better outcomes. The whole thing split on class lines as well, with some commenters defending the idea as a natural extension of daily health monitoring and others calling it a rich-people's biohacking distraction that would do more harm than good to the non-rich.
Only 16 Percent of Americans Think AI Will Have a Positive Impact on Society [comments]
379 points · 457 comments · techcrunch.com · 14h ago
The linked article, from Pew Research, reports that only 16% of Americans see AI as positive for society over the next two decades, with most expecting negative or neutral outcomes and a widespread lack of trust in both government and corporate regulation. The HN thread largely agreed with the poll’s skepticism, arguing that the tech industry has badly overestimated public enthusiasm, and that the backlash has become so normal that using AI in casual chat invites mockery. A major split emerged over whether that distrust is media-driven or, as several people countered, simply a rational response to tech leaders like Altman and Musk openly predicting mass job destruction and societal collapse. Others pushed back on the “too fast” narrative, pointing out that the current AI frenzy looks like yet another speculative bubble—following NFTs and the metaverse—driven by capital that needs a new place to burn, not by genuine demand. A construction-focused commenter derided the idea of replacing skilled trades with AI, noting that “vibecoding” blueprints would lead to expensive, un-fixable real-world mistakes, and that tacit knowledge passed between people over decades simply isn’t in the training data.
AI demands more engineering discipline. Not less [comments]
376 points · 185 comments · charitydotwtf.substack.com · 17h ago
The article argues that as AI code generation becomes cheap and fast, the real engineering work shifts from writing code to specifying, validating, and reasoning about systems—drawing a direct parallel to the earlier shift from pet servers to immutable infrastructure in ops. The thread split hard on this: a vocal group pushed back that the piece itself was meandering and “pointless,” with one person saying the writing felt like “AI slop” and another dismissing the premise that code is no longer the ideal artifact to review, since specs and diagrams remain inherently ambiguous compared to actual code. Others dug into the substance, agreeing with the core analogy but arguing that most tech workers just want a big paycheck and won’t embrace more engineering rigor, while a deeper tangent argued that the real bottleneck isn't code production but evaluation—knowing what the system is supposed to do and whether a new version is correct—which is precisely where human discipline still matters.
RFC 10008: The new HTTP Query Method [comments]
363 points · 152 comments · www.rfc-editor.org · 21h ago
An IETF proposal formalizes a new HTTP `QUERY` method that lets you send a request body with a read-only, idempotent operation—essentially `GET` with a payload, solving the problem of query parameters that are too long for a URL. The thread largely saw this as a clean way to replace the hack of using `POST` for search or GraphQL queries, where the browser can safely retry or cache the result, but several people argued that the real-world caching benefits are overstated since CDNs and proxies would need to start treating request bodies as cache keys, opening up semantic oddities and potential for abuse. A separate crew zeroed in on the fact that RFC numbers just jumped from 9998 to 10008, meaning the long-running prediction market bet on "when will RFC 10,000 land" is now a strange, unresolved mess. The overall takeaway from the room: the spec makes sense for what it targets (large, safe queries), but the practical value of pushing that complexity into generic HTTP infrastructure rather than handling it at the application layer is genuinely up for debate.
Tesco moving 40k server workloads off VMware amid Broadcom's abusive conduct [comments]
294 points · 162 comments · arstechnica.com · 11h ago
The linked article wasn't available to this summarizer; from the discussion, Tesco is reportedly migrating 40,000 server workloads off VMware after Broadcom's aggressive pricing and licensing changes. The HN crowd is deeply divided on which alternative Tesco is actually using—some point to HPE's Morpheus or Red Hat's OpenShift Virtualization, while others champion Proxmox as a viable option at that scale, despite skepticism about managing 40,000 VMs across potentially hundreds of clusters. A significant split emerged over whether Broadcom’s strategy is short-sighted or simply a rational pivot to milk legacy customers as the entire VM market shrinks toward containers and hyperscalers, with several people arguing the long-term brand damage won't matter since Broadcom is now an AI chip company. A few commenters derailed into a heated debate about UK court transparency and whether the Register could even publish details about Tesco's migration, but the core takeaway is that this isn't a unique event—migrations of this size are becoming routine for Red Hat’s team, though many enterprise shops are just paying the insane renewal rates and planning to move slowly.
Why thinking out loud with someone beats thinking alone [comments]
246 points · 105 comments · www.thesignalist.io · 19h ago
The linked article wasn't available to this summarizer; from the discussion, it argues that the act of verbally formulating a half-baked thought into sentences for another person is what actually clarifies thinking, not the listening itself. The HN crowd immediately anchored on the famous anecdote of Einstein thanking a patent office colleague in his relativity paper, which then spiraled into a debate about whether LLMs actually work as rubber ducks—with a vocal contingent insisting that current models are too sycophantic and agreeable, offering no real friction, while another camp countered that even a sycophantic rephrasing forces you to see your own logic more clearly. A deep subthread emerged around whether writing to yourself is better than talking to someone, and the article apparently addressed that directly, which got called out. A surprising tangent: a long discussion about inner monologue structure and whether constructing full sentences in your head is normal or a sign of something else, with several people genuinely confused that others don't do it.
A robot is sprinting towards you. Do you want it running on Claude or Grok? [comments]
239 points · 187 comments · openrouter.ai · 11h ago
The linked article wasn't available to this summarizer; from the discussion, it sounds like someone ran an experiment pitting eleven different LLMs against each other in a 2D battle royale game to see which one performed best, with one model winning 43% of matches and the cheapest beating the most expensive by 27x on cost per win. The HN thread immediately swerved into a meta-debate about whether the article itself was written by AI, with a bunch of people picking apart specific phrases like “I want to be careful here” and em-dash usage as classic LLM tells, while others pushed back saying it’s getting exhausting to see this “critique the writing style instead of the content” pattern on every post. A significant chunk of the thread went full tangent into whether a sprinting robot should bring you a taco and what caliber you’d use to stop a Boston Dynamics dog, with people arguing over 5.56 vs 7.62 vs shotgun slugs for robotic takedowns. The actual split seemed to be between people who think the article is genuinely interesting as a test of model behavior and those who think it’s just another piece of AI-generated slop that’s nearly incoherent and not worth engaging with on its own terms.
MicroUI – A tiny, portable, immediate-mode UI library written in ANSI C [comments]
230 points · 77 comments · github.com · 19h ago
The linked article wasn't available to this summarizer; from the discussion, MicroUI is a tiny, portable immediate-mode UI library in ANSI C that offloads all rendering to the user, requiring only three callback functions for text and rectangle drawing. The HN crowd immediately drew comparisons to Dear ImGui and Nuklear, positioning it as a "bring your own renderer" toolkit for embedding debug UIs into games or low-level graphics pipelines, with one camp arguing its minimalism is a feature for tightly-coupled projects. The biggest split emerged over accessibility: several people dismissed the library outright for lacking any accessibility support, while others pushed back hard, noting that immediate-mode UIs and accessibility are structurally at odds, and that dismissing a 1100-line debug-menu library on those grounds is like calling Unreal Engine a toy project. A related tangent saw someone argue that if accessibility matters, you should just use web tech with semantic HTML and ARIA, which prompted a counter-thread about how even that approach is incomplete and labor-intensive.
Show HN: An 8-bit live gamecast for baseball [comments]
228 points · 121 comments · ribbie.tv · 15h ago
The linked article wasn't available to this summarizer; from the discussion, someone built a site that turns live MLB data into a pixel-art, 8-bit-style baseball gamecast you can watch in the browser. The thread split cleanly into two camps: one group was genuinely charmed by the idea and wanted features like live audio, sound effects, or mini-games during breaks, while another group—several of whom clearly know pixel art—called out the visuals as AI-generated "unfake" pixel art with smeary edges and inconsistent palettes, arguing it doesn't hold up compared to actual NES or TG-16 aesthetics. The creator acknowledged the art was indeed AI-generated and said they'd look into palette reduction and maybe sponsor real sprites later if there's interest, which drew a quick "fine for prototyping" from the critics. A separate tangent made the point that baseball's slow pace is actually what makes this work, unlike faster sports like soccer or basketball where the constant motion would demand a full game engine just to keep up.
Show HN: High-Res Neural Cellular Automata [comments]
193 points · 50 comments · cells2pixels.github.io · 22h ago
The submission is a browser demo and paper on neural cellular automata that now generate high-resolution patterns in real-time by using a lightweight decoder to map cell states to arbitrary pixel coordinates. The thread split sharply on how the models actually work: people who knew the original NCA work from Distill explained that the update rule is just a tiny neural network looking at a 3x3 neighborhood, and that the real novelty is not a texture lookup but a dynamic, self-repairing process where the weights are actually more compact than a JPEG. A running debate broke out over whether the regeneration is truly general or just brittle in the center—one side arguing it collapses if you erase the middle, the other showing it can rebuild from damage if you don't hold the eraser too long, with the consensus settling on "it heals like real tissue, not like a computer undo." Some commenters wanted to paint two competing patterns on the same grid, calling that a missing feature; the author pointed out you can already do it with the transition switch, but acknowledged that the demo's "stamp" mode isn't there yet. A minor but sharp subthread caught the HN spam filter erroneously flagging the submitter's own comment as LLM-slop, with people arguing over whether that was a human flag or an automated kill.
Launch HN: Adam (YC W25) – Open-Source AI CAD [comments]
180 points · 86 comments · adam.new · 15h ago
The post announces CADAM, an open-source web app that uses AI agents to turn natural language prompts or images into parametric 3D models—outputting OpenSCAD code with interactive sliders for tweaking dimensions, all running in the browser. The HN thread was sharply divided: a number of experienced mechanical engineers pushed back hard, arguing the generated models look like "idle doodling" with no manufacturing intent, tolerances, or fasteners, and that saying it made a "V8 engine" is misleading. The submitter countered that this is "AI TinkerCAD" for hobbyists and 3D printing, not professional CAD, and pointed to their separate Fusion plugin for more serious work. Other commenters debated whether the whole text-to-CAD paradigm is viable, given LLMs' still-poor spatial reasoning, or if sketches and iterative visual feedback are more practical. Despite the skepticism, several people said the onboarding and first design experience was genuinely impressive, though the credit-based pricing felt expensive for heavy use.
Anthropic employees accuse Trump administration of targeting them [comments]
176 points · 187 comments · www.nytimes.com · 16h ago
The New York Times article reports that Anthropic employees are accusing the Trump administration of targeting the company, specifically around the release of their "Fable" model after the more powerful "Mythos" was kept internal. The HN thread immediately gets into a dense technical debate about what actually happened—whether Amazon bypassed Fable's safety guardrails or simply got it to do its intended job of fixing code bugs, with one side arguing that "fixing security bugs" is exactly what the guardrails were supposed to shunt to a different model. The bigger split is over whether this is regulation or just a random government grudge: a vocal chunk of the thread insists Anthropic brought this on itself by fearmongering about how dangerous its models were to get regulatory advantage over competitors, and now the administration is using that same rhetoric to single them out for political reasons rather than applying any consistent policy. Others push back hard on that framing, arguing that cheering for the government to "apply the same scrutiny" to other companies is naive—that what's happening here is a clear case of the administration targeting a specific company for not playing ball on weapons systems or national security priorities, which is banana republic stuff regardless of how annoying Anthropic's CEO is.
Storied Colors – A catalogue of named colors [comments]
156 points · 35 comments · storiedcolors.com · 10h ago
The site is a curated catalog of historical pigments, each entry tracing a color's origin, chemistry, and cultural context — from Imperial Yellow's sumptuary laws to the colonial violence behind Indigo. The HN thread quickly split into two camps: one side was captivated by the color stories and book recommendations, while the other immediately flagged the site as "clearly LLM-generated" based on its prose style and design tics, calling it AI slop. The creator waded in to defend the project, explaining they're a technical architect who uses AI to draft entries but then verifies each against primary sources and cuts anything they can't source — a workflow that earned some grudging respect but didn't fully satisfy skeptics. A recurring tension emerged: people who loved the concept felt the LLM presentation undercut the trust needed for a reference work, while others argued that complaining about "vibe-coded" design is just gatekeeping when the content is well-sourced. The thread also surfaced a practical complaint — one entry's claim about hunter-safety data wasn't backed by the cited sources — and a nerdier request for standard pigment codes like PR170 to accompany the hex values.
Local Qwen isn't a worse Opus, it's a different tool [comments]
155 points · 64 comments · blog.alexellis.io · 4h ago
The article by Alex Ellis argues that local models like Qwen 27B aren't "worse" than cloud giants like Opus—they're a fundamentally different tool, better for well-bounded, private tasks like parsing telemetry in airgapped VMs, but prone to catastrophic looping on open-ended agentic work. The HN crowd largely agrees with this framing but pushes back hard on the cost side: multiple commenters point out that even the hardware is brutally expensive (a single RTX 6000 Pro is ~$15k, and a 4090 pulls 350W), and that the real advantage is less about saving money and more about sovereignty and being able to set your own rules. There's a fascinating split between people arguing that benchmarks are worthless because every model requires a completely different prompting style—one person's "vibes" breakdown had everyone nodding along—and a smaller group insisting the real problem is that benchmarks are just gamed by training on the test set, and that we need better evaluation harnesses, not better models. A recurring tangential thread emerges around the idea that LLMs are like instruments, not calculators: you don't just pipe a prompt in, you have to learn to "play" each one differently, and several people share specific hacks (like being overly enthusiastic to Claude so it doesn't get defensive) that confirm the whole thing is more art than science.
Loreline – Tools for writing interactive fiction [comments]
146 points · 18 comments · loreline.app · 11h ago
The linked article is about Loreline, a new open-source tool and language for writing branching narratives, interactive fiction, and game dialogue. While the submission itself is straightforward, the HN discussion mostly spun off into two tangents: first, a running joke that people kept confusing "Loreline" with the totally unrelated "Lore" version control system also on the front page, with a handful of comments playfully lamenting the confusion; second, a series of comparisons to established tools like Inkle's Ink, ChoiceScript, and the legendary Inform 7, with one defender arguing that Inform 7's natural-language prose approach is actually excellent for its target audience of writers, despite how much it horrifies programmers. A minority of commenters asked about specific features like out-of-the-box web export targets and found the docs helpful, but the main energy of the thread was people either getting the name wrong or defending the weirdness of prose-based IF languages.
Taxonomy of the Occlupanida (parasitoids on bread bag tags) [comments]
140 points · 33 comments · www.horg.com · 8h ago
The linked article is a deadpan synthetic taxonomy of those plastic bread bag clips, complete with absurd Latin names and families like “Acutignathidae” and “Megapanis gregjubyi.” A fair chunk of the thread took the “parasitoid” title literally, thinking the site was about actual biological parasites, and had to reorient once they realized it was a joke—though a few held out hope for someone actually culturing bread tags. The discussion split between people who treat these clips as serious, functional objects (with ecological niches and even mating behavior) and those who see them as a pointless, annoying invention that breaks or mauls the bag, with Europeans particularly adamant that they’d never seen the things. Others shared genuine memories of museum exhibits with thousands of bread tags on display or of using the clips as temporary sandal repairs and makeshift ninja stars as kids—a reminder that this is a bit of Americana that never quite took root elsewhere.
The Competitive Moat That AI Can't Replicate [comments]
136 points · 114 comments · ghostinthedata.info · 14h ago
The article is a long-form argument that human connection is AI's un-replicable moat, using examples like a restaurant that switched to online booking but kept staff as mini-concierges, and Danny Meyer's distinction between hospitality and service. The HN thread immediately broke into two camps: people who agreed with the thesis and people who were convinced the post itself was AI-generated slop, pointing out the telltale sentence fragments, obsessive three-item lists, and "not X, but Y" structures. A running sub-thread debated whether an AI-detection tool called Pangram is actually reliable or just HN's astrology, and someone pointed out the irony of an AI-written post arguing for the importance of human connection. The deeper pushback came from people who said they *don't* want a "connection" with a business — they want transactional interactions that actually work, and that AI chatbots that simulate service without providing it are making things worse, not better.
French physicist and media star loses doctorate after plagiarism investigation [comments]
132 points · 138 comments · www.science.org · 16h ago
A French physicist and media personality, Étienne Klein, had his doctorate formally revoked after an investigation uncovered plagiarism on 20% of the thesis pages. The thread focused heavily on the mechanics of the detection, with many arguing that modern AI tools—specifically large language models—made it trivially easy to surface the kind of lightly rephrased passages that were nearly impossible to catch when he wrote it 30 years ago, but the consensus was split on whether that makes the crime worse or just changes how it gets caught. A significant faction pushed back against the idea that this was a serious academic transgression, arguing that a PhD thesis is more of an intellectual exercise and a formality than a piece of original, publishable research, and that Klein’s real crime was not plagiarism but rather failing to acknowledge that he was a popularizer, not a researcher. Others, however, pointed to his wildly disingenuous defenses—including that he had memorized books by reading them aloud and that he was being attacked by anti-science forces—as evidence that he knew exactly what he was doing, and the discussion ultimately landed on the broader question of whether the entire academic system, and the thesis itself, has any real value if nobody on the committee actually reads it.
ICE Appears to Be Buying Immigrants' Tax Identifiers from a Data Broker [comments]
121 points · 55 comments · www.404media.co · 17h ago
The linked article wasn't available to this summarizer; from the discussion, it's about ICE reportedly buying immigrants' tax identifiers, specifically ITINs, from data brokers. The thread immediately pounced on the core legal question: if the government can't directly access this data without a warrant, can it legally outsource that surveillance to a private company? A strong contingent says no, arguing this is a blatant loophole that lets ICE skirt the Fourth Amendment by purchasing what it can't compel, with one commenter pointing to the SCOTUS ruling on cell tower data as the obvious precedent to extend. Others pushed back hard against that analogy, saying the government already buys tons of data from brokers for all sorts of purposes and that "the government should not be able to outsource things they cannot do" is a silly principle when applied to something as mundane as socks for the army—though the thread itself admitted the original phrasing was poorly worded. A separate, darker vein of discussion ran on the premise that the real scandal is the tax system itself, which funnels billions from undocumented workers into unclaimable revenue while leaving them vulnerable, and that ICE buying this data just turns a massive money machine for private brokers into an even worse surveillance apparatus.
Clojure Hosted on Go [comments]
121 points · 14 comments · github.com · 8h ago
The submission is about Glojure, a new interpreter that lets you run Clojure code directly on top of Go, giving Go programs access to Clojure's expressiveness and Lispy tooling. The thread was split between excitement about how this fills a real gap — people want Clojure's REPL and macros in the Go ecosystem, especially for scripting and embedding — and immediate skepticism about performance, since it's a tree-walking interpreter and not JIT-compiled, which will be slow for anything compute-heavy. A few people pointed to competing projects like Jank (which targets C++ with an LLVM JIT) and Lisette (typed FP on Go), but the consensus among those who've tried the alternatives was that Glojure is the only one with proper, full interop with Go's standard library, so it actually works for real integration. There was also some noise in the comments about the project's maintenance history bouncing between forks, but the developer has since consolidated everything back into the upstream repo and is pushing forward under a Clojurists Together grant.
How Madrid built its metro cheaply (2024) [comments]
120 points · 63 comments · worksinprogress.co · 12h ago
The article explains how Madrid tripled its metro in 12 years for a fraction of what U.S. and U.K. projects cost, pointing to specific tactics like standardized station designs, 24-hour permitting, and keeping a core in-house engineering team. The thread mostly dug into whether those tactics could work in the U.S., and the consensus was basically "no, not without a totally different political and economic setup." People split sharply on why U.S. labor is so much more expensive—some blamed housing costs, but a solid pushback argued it's just that the U.S. economy is so much more productive and its unemployment so much lower that construction wages here can't help but dwarf Spain's. A few commenters also zeroed in on the in-house engineering point as the real takeaway, contrasting it with how California projects like the Caltrain Downtown Extension get endlessly derailed by politicians bolting on unrelated demands and wasting years on environmental reviews.
SteamOS Linux 3.8 released as stable [comments]
116 points · 20 comments · store.steampowered.com · 3h ago
This is about the release of SteamOS 3.8, which updates the underlying Arch Linux base from roughly 1.5-year-old packages to 1-year-old ones, bringing a big KDE Plasma jump from 6.2.5 to 6.4.3 and adding native Wayland support. The thread immediately latched onto the fact that the release is timed a full year after KDE 6.4 shipped, with one person who's been daily-driving an older build on a monitor setup saying they're finally willing to wipe their desktop for it. There's a split on whether SteamOS is usable as a general-purpose desktop distro: some point to Distrobox as the key enabler, while others flag that it doesn't work with NVIDIA GPUs and suggest CachyOS or Bazzite instead. A tangent pushed for Blackmagic Design to build DaVinci Resolve on SteamOS rather than Rocky Linux, though the counter from someone in the VFX space is that a rolling distro with no pinned glibc version is a non-starter for industry toolchains.
With Wall Street’s help, you’re about to be forced to buy stock in SpaceX [comments]
99 points · 72 comments · paulkrugman.substack.com · 11h ago
Paul Krugman’s substack piece argues that Elon Musk is a “human Ponzi scheme” whose trillion-dollar valuation is propped up by investor faith in his genius, not actual product delivery, and that the recent SpaceX IPO—with a $1.77 trillion valuation on just $18.7 billion in revenue—is the clearest example yet, especially since index funds changed their rules to force ordinary Americans to buy in. The thread immediately split into the familiar “what about EVs and Starlink” defense versus Krugman’s point that those real achievements aren’t what made Musk the richest man in the world, with several people quoting the article back to sarcastic detractors who clearly only read the headline. A major pushback came on the “Human Ponzi” label itself, with people arguing it’s a weak comparison because no one is promised a guaranteed return, and that calling any overhyped stock a Ponzi scheme cheapens the term—though one reply countered that the whole point is index funds are now forcing you to speculate, so “just don’t buy” isn’t simple anymore. The thread also got into the weeds on whether X is really an “extreme-right Nazi-friendly cesspool,” with one side saying the article is being honest about free speech being an existential threat while the other side pointed to actual data showing Twitter’s top engagement now comes from the right, and a sub-comment noting the difference between “right-leaning” and “Nazi-friendly.” Several people took shots at the idea that Musk has little technical understanding, but the responses pointed to a link that supposedly debunks the trope, while others countered that even if he’s not a real engineer, his real skill is the confidence to sell impossible-seeming plans to retail investors who don’t care if the prophecy comes true as long as the stock price does.
ChatGPT's image generator can be manipulated to produce violent, sexual content [comments]
92 points · 132 comments · mindgard.ai · 7h ago
The article recounts how Mindgard researchers found that ChatGPT’s image generator can be tricked into producing explicit sexual violence and gore simply by claiming a fake image was “already approved” or by repeating a viral prompt with minor wording changes — and that the model consistently chose the worst possible output whenever given latitude. The HN thread quickly turned into a familiar but sharp debate about whether LLMs are fundamentally unfixable: one side argued that because user input and system prompts are mashed together in the same attention matrices, prompt injection and content bypass are architectural inevitabilities, like an OS that marks all memory executable. The counterargument was that this is just a cat-and-mouse game of mitigations, and that simple post-generation classifier models could at least catch nudity and gore — but others shot back that even that failed with Fable, which couldn’t distinguish biology from safety concerns. A vocal minority went further, accusing OpenAI and other lab leaders of knowingly including CSAM in training data and calling for congressional hearings, while a running joke referenced Microsoft Tay as the prophet no one listened to. The thread never really resolved the split; it just doubled down on the deep disagreement over whether the problem is a bug or a feature of the architecture.
Generated 2026-06-18 08:30 UTC
Generated by Sauron from Hacker News discussions and linked articles.