HN Brief: 2026-08-29
Today's HN was split between practical infrastructure gripes and geopolitically charged tech policy. The big threads circled three themes: first, a collective frustration with software that's *almost* good but broken by deliberate design choices — keyboard navigation that's been hidden in accessibility settings, apps that redirect you to the web for basic features, and update cycles that ship "improvements" that make things worse. Second, the US Treasury's designation of the Italian activist-host collective Autistici/Inventati as a terrorist entity dominated the political side, with threads debating whether "anti-fascist" infrastructure is fair game for sanctions and what that means for the open web. Third, model releases and AI infrastructure kept appearing, from GLM-5.3's open-weight release to OpenAI cutting off Cursor after SpaceX bought it, alongside a growing sense that LLM tooling is fragmenting faster than anyone can keep up with.
The threads worth clicking into: "GUIs should be fully keyboard-driven" for the sharp takedown of Electron and Apple's buried accessibility settings; "Get your Windows license refund" because the debate over whether OEM Windows costs $200 or $2 reveals how little we actually know about PC economics; "U.S. sanctions against the A/I Collective" for the uncomfortable question of whether requiring users to share anti-fascist principles makes you a terrorist or just selective; "I accidentally turned LLM memory into program analysis" for the Datalog-powered fact-tracking approach that might actually fix LLM hallucination for long research sessions; and "Verschlimmbesserung" because the thread's hunt for the perfect word for updates that make things worse is cathartic and genuinely linguistically interesting.
GUIs should be fully keyboard-driven [comments]
757 points · 382 comments · ckardaris.com · 16h ago
The article argues that GUIs should be fully keyboard-driven, pushing back against the idea that this is a reason to prefer TUIs over graphical apps. The thread largely agrees that full keyboard navigation is a worthy goal, with much of the discussion shifting to a sharp critique of modern software’s failure to deliver it, especially on macOS and the web. A major thread blames popular frameworks like Electron for dropping the ball on tab order and shortcut consistency, while others point out that Apple deliberately buries full keyboard navigation in accessibility settings, and that Microsoft’s older Windows components were far better at it. The conversation also splits on whether discoverability is the real problem—one side argues for on-screen hints and standardized shortcuts (like vim bindings), while another insists the ideal is a universal command palette, à la JetBrains IDEs, that lets you do anything from the keyboard without memorizing chords.
Get your Windows license refund [comments]
679 points · 281 comments · en.refund4freedom.org · 18h ago
This is a guide and advocacy site pushing for consumers to demand refunds on Windows licenses bundled with new laptops they never intend to use, complete with real legal case histories from Italy showing refunds ranging from €40 to €129. The thread quickly got into a debate over OEM pricing, with several people pushing back hard on the idea that manufacturers pay anywhere close to the retail license cost—pointing out that $200 Windows laptops exist, so Microsoft must be selling licenses for nearly nothing to keep Asus and others from switching to Linux or Chromebooks. Others dove into the practical reality of doing this, with one person noting they kept Windows purely for resale value since "the year of the Linux desktop" hasn't actually arrived, while another reported reviving a 15-year-old Thinkpad with Linux Mint and an SSD and finding it far snappier than it ever was with Windows. A side argument broke out about whether Linux genuinely runs well on modern hardware, specifically around web browser performance, codec support, and the total lack of support for Apple Silicon Macs, with one person arguing that performance degradation is purely a software problem and that JavaScript bloat is the real culprit across any OS.
GLM-5.3 is now open-weight [comments]
678 points · 226 comments · huggingface.co · 16h ago
The linked article is the Hugging Face release of GLM-5.3, an open-weight model that improves on its predecessor entirely through post-training, claiming massive gains in complex coding and a newly emergent "cyber capability" for vulnerability discovery and exploitation. The HN thread largely agrees this is a serious, top-tier open model, with multiple people directly comparing its feel and quality to Anthropic's Opus 4.8 and calling it a current sweet spot for local or API use, especially since it avoids the "load-bearing honesty" and geopolitical restrictions of US labs. A huge slice of the conversation focused on practical deployment: people are already running it via DeepInfra, Ollama Cloud, and other providers, though the GLM-5.3-Flash variant drew sharp splits over its pricing (some found it cheaper, others more expensive than DeepSeek Flash) and notably inconsistent, sometimes very slow inference speeds. The discussion also dug into model size speculation, noting the full-precision version is around 770GB and that this Chinese lab is operating under far tighter hardware constraints than Western counterparts, making the results more impressive rather than embarrassing.
“It works better in the app” [comments]
650 points · 443 comments · shkspr.mobi · 19h ago
The article is a frustrated rant from a developer trying to add an external calendar URL to a Google Pixel phone, only to find the Google Calendar app has no way to do it—forcing a desktop-mode browser workaround, and sparking a broader complaint about half-finished apps that push you to the web for anything outside their narrow scope. The HN discussion largely agreed with the premise but split on the root cause: some argued it’s not incompetence but deliberate friction to lock you into Google’s ecosystem, while others pinned it on a corporate culture that rewards launching shiny new features and then abandoning the boring maintenance work. A major tangent erupted over why companies force you into apps at all—the consensus was that it’s about tracking, notifications, and data collection, not user experience—with many commenters proudly refusing to install social media or banking apps, using browser workarounds or PWAs instead. The thread also bickered over whether native apps still offer anything the web can’t do well, with defenders pointing to NFC, offline reliability, and real-time performance, while detractors countered that modern browsers handle video, sensors, and even Half-Life 2 just fine.
Htmx 4.0 [comments]
632 points · 155 comments · four.htmx.org · 18h ago
Htmx 4.0 is out after eight months of work, migrating internally from XMLHttpRequest to `fetch()`, making attribute inheritance opt-in with `:inherited`, cleaning up event naming, and dropping localStorage-based history caching in favor of server re-fetches. The HN thread quickly homed in on the version numbering joke — the team promised never to do a v3, so they skipped straight to v4, which got both applause and a barb about "confident solutions based on half understanding." A real split emerged over LLMs: some argued that since AI can now generate plain JavaScript easily, htmx's simplicity advantage is eroding, while others countered that htmx's minimal surface area is *more* useful with LLMs because the shorter context window means better code generation. Plenty of people chimed in with their preferred stacks (Go+SQLite, Django+Postgres, Rust+Axum) and shared that htmx paired with Alpine.js and Claude is their current go-to for fast, maintainable prototypes.
U.S. sanctions against the A/I Collective [comments]
569 points · 556 comments · www.inventati.org · 19h ago
The US government sanctioned Autistici/Inventati, an Italian volunteer-run collective that has provided free email, web hosting, and encrypted communication tools to activists and grassroots groups since 2001. The Hacker News thread immediately split on whether A/I is a neutral infrastructure provider or an activist group that vets users for ideological alignment—the collective openly requires users to share principles like anti-fascism, anti-racism, and anti-militarism, which several people pointed out is clearly stated in their manifesto. A big chunk of the discussion zeroed in on the State Department’s framing that “anti-fascism” qualifies as a terrorist ideology, especially since the Trump administration had already designated Antifa as an organization by fiat, and people argued that resisting fascism is a centrist position, not radical left. Others pushed back that being selective about who gets free services means you can’t claim the “we’re just pipes” defense when users break laws, though the counter was that refusing service to fascists is the normal right thing to do and doesn’t make you an activist gun seller. There was also confusion over what A/I actually does—some found the website vague—until multiple commenters clarified they host noblogs, run mailing lists, and provide VPNs specifically for anarchist, anti-fascist, and anti-imperialist movements.
Inception-style curved map for turn-by-turn directions [comments]
487 points · 161 comments · www.orbify.eu · 19h ago
The submission is a demo of a curved, "Inception-style" map for turn-by-turn navigation that blends a driver's perspective with an overhead view. HN was immediately split: many praised it as a genuinely useful way to see turn angles and lane choices at a glance, while a much larger group reported feeling nauseous or dizzy even at the mildest curvature, with the creator acknowledging the effect needs toning down and promising configurable sliders. Several commenters pointed out that this concept has extensive prior art—most notably Berg London's "Here and There" posters from 2009 and a Disney multiperspective camera paper from 1997—and flagged that an existing patent already covers the exact use case until 2030, which may limit any patent claims. A recurring critique was that during the turn itself, the view provides no information about what comes immediately after, making consecutive turns hard to follow, and some argued that a well-zoomed, motion-oriented top-down map already solves the same problem without the motion sickness.
Our decision on Cursor following its acquisition by SpaceX [comments]
465 points · 234 comments · openai.com · 6h ago
OpenAI announced it will stop providing future models to the AI coding assistant Cursor, which SpaceX just acquired, though existing models remain available for now. The HN thread largely dismissed this as a minor blow, with multiple people claiming only around 5% of Cursor users actually rely on OpenAI models—most use Cursor’s own Grok-powered models or bring their own Claude/Anthropic keys. The real split in the comments is between those who see this as OpenAI circling wagons ahead of its Astra model launch and those who think Cursor’s entire reselling-others-APIs business model was always on borrowed time, with several arguing Claude Code has already surpassed Cursor for serious agentic coding anyway. A few people noted the irony of Elon Musk’s SpaceX buying a tool that heavily depended on the company he constantly sues, and speculated whether Anthropic might follow suit given their own anti-distillation stance.
Just the rumour of a bug is enough to find an exploit these days [comments]
303 points · 107 comments · anil.recoil.org · 16h ago
The article details a real-world case where the author patched a path traversal bug in OCaml's cohttp library, only to see probes hitting his server within minutes of opening the public pull request. HN’s rclone maintainer chimes in to confirm this is their life now—after a decade of 20 security disclosures total, they got over 40 in the last month alone, and GitHub’s CVE assignment has slowed from days to weeks because human triage is swamped. Several top comments push back on the novelty claim, arguing that mining exploits from patch diffs and commit messages is as old as vulnerability research, but LLMs have “democratized” it to the point where the entire chain—reading the patch, building the exploit, scanning the internet—can run automatically in minutes instead of days. A pointed sub-thread argues the real bottleneck isn’t AI but managerial will: developers can find and fix bugs faster than ever, but bosses still prioritize velocity over quality, and one commenter describes getting hauled into a multi-meeting circus for fixing a datetime bug that sent false alerts. The consensus splits between those who see this as a systemic process failure in open-source packaging and release infrastructure, and those who think the deeper rot is that nobody in power actually wants to pay for the grind of making software not break.
Hilariously fast volume computation with the divergence theorem (2018) [comments]
256 points · 66 comments · alyssarosenzweig.ca · 23h ago
This is a blog post from 2018 that walks through deriving a fast O(n) algorithm for computing the volume of a closed 3D mesh using the divergence theorem, boiling it down to a simple loop over triangles with just a few arithmetic operations per triangle. The HN discussion quickly zeroed in on the fact that this isn't new at all—people recognized it as the 3D analog of the shoelace formula (or surveyor's formula) for area, with several pointing to prior art from the 1970s and 1980s. A few commenters pushed back on the author's characterization of "naive algorithms" as rendering and sampling, arguing that the obvious geometric approach (summing signed tetrahedron volumes to the origin) is just as straightforward and equally O(n). There was also a split over whether the calculus derivation is necessary or if pure geometric intuition suffices, though most agreed the write-up is a fun, clear exercise that doesn't pretend to be novel.
EPA says power for data centers can sidestep pollution laws [comments]
244 points · 251 comments · www.epa.gov · 18h ago
The EPA issued guidance declaring that data center power plants not connected to the public grid are exempt from Clean Air Act acid rain regulations, effectively letting them burn fuel without pollution controls as long as they stay "islanded." HN immediately jumped on the wordplay—the P in EPA now stands for "Pollution," "Profit," or "Privatization," depending on whom you asked—and a long, detailed comment tore the agency's legal rationale apart, showing it hinges on a 1990 survey form's obsolete definition of "utility" that the EIA itself abandoned. The debate split between people arguing this is a straightforward reading of the letter of the law versus those saying it flagrantly violates the spirit, with one responder who grew up in the 1970s noting acid rain was turning surface water to pH 1 before regulations kicked in. Several observers warned this is pouring gasoline on the already blazing bipartisan hatred of data centers, while others pointed out the real driver is timeline—grid interconnection takes years, and onsite generation can be built alongside the data center, though it runs on dirtier single-cycle gas turbines. A few commenters also flagged that the guidance document itself misquotes its own source, writing "primary" instead of "primarily," which they called an amateurish anti-signal for a document making a major regulatory carveout.
Boot a Virtual iPhone via Apple's Virtualization.framework [comments]
239 points · 69 comments · github.com · 9h ago
This is a project that lets you boot a full virtual iPhone on Apple Silicon Macs using Apple's own Virtualization.framework, essentially running a real iOS kernel and userspace rather than the stripped-down simulator. The HN crowd immediately got into the weeds comparing it to Xcode's Simulator, with the consensus being that this is fundamentally different—the Simulator runs iOS frameworks as native macOS processes sharing your kernel, while this boots a complete, separate iOS kernel in a VM, which matters for testing things like network extensions, VPNs, camera APIs, or jailbroken environments that the Simulator simply can't do. A big thread emerged around the requirement to partially disable SIP and AMFI, which several people pointed out makes this a non-starter for corporate machines or MDM-controlled devices, though others pushed back that you can buy a separate research machine or use a more surgical allowlist approach with `amfidont`. There was also a fascinating rabbit hole about why picking Japan or the EU as your region during iOS setup breaks the VM—Japan apparently has Felica chip expectations and a side-button rebinding entitlement, while the EU triggers third-party app store availability checks that the VM can't spoof, leading to speculation about spoofing cell tower data or billing addresses to bypass it.
Chicken products recalled in five states due to “false marks of inspection” [comments]
194 points · 161 comments · www.thehealthy.com · 12h ago
The USDA recalled 24,900 pounds of frozen Buffalo chicken rangoon and mozzarella sticks from a Boston distributor because the products carried fake inspection marks, part of a growing pattern of enforcement against possibly unexamined imported meat. The discussion immediately turned political, with one camp blaming recent DOGE cuts to the FDA and USDA for gutting food safety—pointing to only 400 inspectors covering 36,000 facilities—while others countered that major outbreaks like Boar's Head and McDonald's E. coli both happened in 2024, before those cuts, and that fewer recalls under Trump might just mean less testing, not safer food. A faction of commenters argued this specific recall actually shows the system catching fraud correctly, not failing, and dismissed the panic as narrative-driven. A few people went off on the genuinely surreal product names—"Buffalo chicken rangoon" and "Benedetto's Buffalo chicken mozzarella sticks"—as a distraction from the serious question of whether the U.S. food supply is getting riskier or just getting more attention.
Migrating to HTTPX2 [comments]
193 points · 82 comments · github.com · 20h ago
The OpenAI Python SDK has switched its default HTTP client from httpx to HTTPX2, a fork maintained by the same team behind Pydantic, and the migration document lays out how to update custom transports, auth handlers, and mocking libraries. The HN discussion quickly turned into a referendum on the original HTTPX project, whose maintainer locked down issues and discussions citing an "absurdly skewed gender representation" in the online community, then moved development to a private repository — effectively abandoning the library and leaving projects like MkDocs in limbo. Several commenters pointed out that Anthropic made the identical switch shortly after OpenAI, and that the real driver is less about technical superiority and more about HTTPX being a zombie project that can't even reach a 1.0 release without breaking everything. The shift from certifi to the operating system's TLS trust store drew mixed reactions: some called it a welcome fix for corporate proxies and container cert headaches, while others warned it will silently break deployments on minimal container images that lack system CA bundles. A few commenters grumbled about yet another HTTP library fork polluting the Python ecosystem, but the consensus was that the Pydantic-backed fork (with involvement from the new Starlette maintainer) is the de facto future and the old HTTPX is effectively dead.
I used AWS cognito for a startup. I wouldn't do it again [comments]
176 points · 124 comments · joshkaramuth.com · 18h ago
The post is a detailed account from a developer who regrets choosing AWS Cognito for a startup's authentication, citing terrible documentation, breaking changes between Amplify versions that forced a full rewrite, impossible-to-fix configuration mistakes that require deleting the entire user pool, and a painful lack of local development support. The thread largely agrees with the author, with many saying this is the standard experience for all of AWS's documentation—exhaustive but exhausting, filled with deprecated examples and information that’s impossible to find again. A significant split emerged around whether the article itself was AI-generated slop, with one faction calling out specific phrases as telltale signs and another arguing the underlying critique is valid regardless of how it was written. People piled on with their own Cognito horror stories, especially the inability to change email case-sensitivity or export password hashes without forcing users to reset, which several commenters pointed out makes migration to any other provider a nightmare. The consensus was clear: Cognito is cheap and integrates with AWS, but you’ll pay for it in engineering hours, and if you ever want to leave, you’re stuck.
Stopping the smart TV from being used against you [comments]
174 points · 139 comments · www.s-config.com · 11h ago
The article is a furious rant about how LG smart TVs can install unwanted software like McAfee onto a Windows laptop via the HDMI cable's EDID protocol, framing it as malware that violates user trust. The HN crowd quickly split on the delivery: several people found the author's prose—including the use of “rape” as a metaphor and the breathless, stream-of-consciousness style—exhausting and distracting from a real grievance. The technical pushback focused on EDID, with commenters pointing out that the protocol is a harmless standards-based handshake for display identification, not a vector for malware, and that the real problem is Windows letting LG's driver installer run wild. The discussion pivoted hard into practical defenses: the strongest consensus was to never connect a smart TV to the internet at all, with a deep sidebar war breaking out over whether your ISP's cable modem is secretly hosting public hotspots for other customers, and another thread arguing that if you want to stream, just buy a dumb display and use a separate Linux box—though someone immediately undercut that by noting you still can't stream Netflix at native resolution on Linux.
State of the Map 2026 [comments]
159 points · 54 comments · 2026.stateofthemap.org · 18h ago
The announcement is for State of the Map 2026, the annual OpenStreetMap conference, happening near Paris at the end of August 2026. The HN thread quickly swerved past logistics into a political landmine: does OSM comply with the US executive order renaming the Gulf of Mexico and Lake Ontario? The consensus is that OSM handles it with nuance — the default `name` tag sticks with what locals actually call it (Gulf of Mexico), while `official_name:en-US` captures the new "Gulf of America" for search and geocoding, with most renderers ignoring the official name outright. A bunch of people pointed out that the president only has authority over federal maps, and that Google and Apple are geofencing the rename to US viewers, which some called a sad kowtow and others just pragmatic compliance. A separate long thread dove into a quirk in the `mkgmap` tool that transliterates Japanese POI names into Chinese pinyin when English fallbacks are missing, exposing how political border-drawing and language assumptions lurk in even the most mundane OSM data pipelines.
Autistici/Inventati's main .org domain goes dark after US terrorism designation [comments]
140 points · 97 comments · mastodon.bida.im · 17h ago
The linked article wasn't available to this summarizer; from the discussion, Autistici/Inventati, an Italian collective providing internet and hosting services to activist and social movement groups, has had its main .org domain taken offline after being designated a terrorist entity by the U.S. Treasury for allegedly supplying digital architecture to "Antifa cells and other violent far-left extremists." Hacker News immediately zeroed in on the Treasury's framing, which explicitly lists "Violent Left-Wing Extremists, including Anarchists and Anti-Fascist" as a major terror threat while omitting any mention of right-wing extremism, despite the fact that right-wing attacks have accounted for the majority of U.S. terrorist incidents since the 1990s. Several deep dives argue that "Antifa" is not a coherent organization but a loose label, and that designating it as a terrorist group is a political move that conflates militant anti-fascism with far-left violence, drawing direct parallels to how the Trump administration declared Antifa a domestic terrorist organization in 2025. A significant split emerged between those who see this as a dangerous expansion of state repression targeting legitimate political dissent and those who counter that far-left and Islamist violence pose a far greater threat, with one commenter pointing out that the Treasury's categories conveniently sidestep foreign-linked right-wing networks like The Base and Atomwaffen, which operate internationally with impunity.
EasyEffects can improve laptop speaker sound quality [comments]
138 points · 53 comments · www.osnews.com · 16h ago
The article argues that EasyEffects, a Pipewire-based audio processing tool for Linux, should be bundled by default in every distribution and desktop environment because it can dramatically improve the tinny, hollow sound of laptop speakers using community-shared presets. The HN thread largely agreed that software tuning is essential—most Windows laptops already ship with something like Dolby Atmos for the same reason—but the discussion quickly pivoted to a debate over whether aiming for a "flat" frequency response is even the right goal, with some arguing that flat is the objective reference for mastering and others insisting that subjective bass-boosted or "warm" presets sound better to most human ears. Several people shared concrete success stories, especially Framework laptop owners who reported their machines went from "absolute trash" to "plenty fine" with specific EasyEffects presets, and a few pointed to a Python script that extracts Dolby Atmos tunings from Windows drivers and converts them for use in EasyEffects. A notable tangent argued that while the idea is sound, integrating it into the default system volume control with loudness compensation would be far more impactful than just shipping a preset, and one person wished for an auto-tune feature that uses the laptop's own microphone to calibrate the EQ automatically.
Smaller reactors bring nuclear power closer to fulfilling its promise [comments]
136 points · 192 comments · www.nature.com · 19h ago
The Nature piece surveys the push for small modular reactors (SMRs) as a potentially cheaper, safer, and more flexible alternative to conventional nuclear plants, highlighting designs like Kairos Power's TRISO-fueled pebble-bed reactor and TerraPower's sodium-cooled Natrium. The HN crowd immediately pounced on the word "should" in "SMRs should be cheaper," arguing that nuclear economics have always collapsed under cost overruns and that the industry is betting on unrealistic learning curves. A major split emerged between skeptics who see solar-plus-battery as already winning on cost and speed, and defenders who insist that firm, weather-independent power is necessary to avoid massive overbuild and to handle multi-day "Dunkelflaute" events in northern winters. The renewable camp fired back that you only need gas peakers for the rare gaps, and that 95% decarbonization beats chasing the fantastically expensive last few percent with nuclear, while the nuclear side pointed to land-use constraints and the reality that SMRs still need HALEU fuel that doesn't yet exist at scale. A secondary thread worried about physical security of distributed smaller reactors and cynicism that these are really just grifts for Big Tech's AI data centers, though others countered that every energy source has risks and nuclear remains statistically the safest.
Interactive Warhammer 40k Galaxy Map [comments]
134 points · 43 comments · cartographia40k.com · 23h ago
The linked article wasn't available to this summarizer; from the discussion, it's an interactive 3D map of the Warhammer 40k galaxy, letting you pan and zoom around named worlds, factions, and notable locations. The thread quickly split into a handful of camps: people who love the thing as a reference tool, people whose browsers immediately choke when they toggle the route lines on, and a surprisingly intense linguistic debate about whether the Latin in the map's labels is correct or whether it's supposed to be deliberately mangled "High Gothic." That debate then detonated into a full-blown argument about whether the entire 40k setting was originally a joke, a satire of 1980s Thatcherite England, or just drunk nerds riffing—with one side insisting the Emperor is a dead-Jesus allegory for the Church of England and the other side calling that interpretation way oversold. A smaller, more pragmatic faction just wants source citations per planet so they can figure out which book to read next, and one person jokingly demanded the Inquisition terminate the "Abominable Intelligence" that generated the map.
An investigation into the state of corvid–human relations [comments]
128 points · 77 comments · www.audubon.org · 19h ago
The Audubon article explores whether the relationship between humans and crows qualifies as genuine friendship, diving into the science of corvid recognition, gift-giving, and the emotional pull of feeding them. The HN thread largely ignored the article's philosophical questions about friendship and instead pivoted hard into crow war stories: people swapping tales of being personally targeted by poop strikes, witnessing crows bullying swans for sport, and the legendary grudges crows hold and pass down through generations. A significant split emerged between the Corvid Defense Force, who pointed out that Australian "attacks" are actually from magpies (which aren't even corvids), and the lawn-owners who argued that crows destroying turf for grubs is a real economic threat that justifies lethal control. The practical takeaway from the thread is that you can win most crows over with peanuts, but if you get on their bad side, you're making enemies for life—and for your children's lives.
Verschlimmbesserung: The Word Your Software Updates Need [comments]
127 points · 85 comments · geekyschmidt.com · 17h ago
The article introduces the German word "Verschlimmbesserung" — an attempted improvement that makes things worse — as a perfect label for modern software updates that break workflows, move buttons, or ship for the sake of shipping. Hacker News immediately debated whether English already covers this ground, with one camp arguing "enshittification" is close enough, though others pushed back hard, insisting that enshittification implies intentional malice while Verschlimmbesserung is about misguided good intentions or metrics-driven incompetence. Native German speakers split on whether the word is still common, with some calling it everyday vocabulary and others saying they haven't heard it in years. The thread also surfaced competing coinages like "worseprovement," "misprovement," and "updowngrade," and one linguistically-inclined commenter broke the grammar down in detail. A recurring point was that the real problem isn't the word itself but the incentive structures that reward shipping churn over stability — and that Office 2003 was a better product than most of what we get today.
9th Circuit sides with states in Kalshi gambling fight [comments]
122 points · 97 comments · azmirror.com · 8h ago
The linked article reports that the 9th Circuit Court of Appeals ruled that sports betting doesn’t qualify as a “swap” under federal commodities law, meaning states like Arizona can resume prosecuting companies like Kalshi for running unregulated prediction markets. The court’s unanimous decision by three Trump appointees explicitly rejected a broader reading of the Commodity Exchange Act, saying Congress didn’t intend to bulldoze state gambling regulations when it amended the law — a direct split from a prior 3rd Circuit ruling that had sided with Kalshi. The thread largely agrees that this outcome was obvious, with several people arguing the law’s slow, deliberate pace ultimately prevents the kind of bad rulings that would immediately stick, though a counterpoint notes that same slowness lets administrations like Trump’s do lasting damage before any check lands. A major side-conversation erupts over whether gambling itself should just be allowed but heavily taxed and regulated like tobacco, versus critics who insist that prediction markets create dangerous incentives to manipulate real-world events, with one commenter pointing to reports of war intelligence being distorted because someone bet on frontline movements.
HTTPX2 – A next-generation HTTP client for Python [comments]
119 points · 38 comments · github.com · 20h ago
Pydantic has forked the popular but slow-moving HTTPX library as HTTPX2, taking over maintenance and promising timely security updates for a library that sits in the critical path of many production systems. The HN thread immediately swerved into a naming debate, with several people arguing "HTTPX2" is confusing because it sounds like the HTTP/2 protocol rather than a continuation of the HTTPX project, and suggesting alternatives like "httpx-ng" or "httpy." A vocal contingent pushed back hard on the whole premise, saying HTTPX already performs significantly worse than aiohttp at scale and that anyone serious about performance should be using niquests or pyreqwest instead, with some commenters calling niquests clearly superior despite it being a one-person project. The thread also dug into the broader Python HTTP library proliferation mess, explaining that requests is feature-frozen without async support, httpx's original maintainer closed off all outside contributions for personal reasons, and the ecosystem keeps forking rather than consolidating because HTTP is deceptively complex once you get into async, connection pooling, and newer protocol support.
I accidentally turned LLM memory into program analysis [comments]
113 points · 24 comments · pwning.systems · 8h ago
The author built a system called Lemmalog that uses a Datalog engine to maintain an LLM's state during long vulnerability research sessions, turning observations into structured facts with tracked dependencies so that when a fact is disproven, all conclusions derived from it are automatically invalidated. The HN thread immediately recognized this as a revival of old-school symbolic AI, with several people drawing direct parallels to CodeQL, Prolog, and even Cyc, arguing that this is exactly the kind of hybrid approach needed to ground LLM outputs in deterministic reasoning. Others pushed back on the practicality, noting that the hard part isn't the fixed-point computation but reliably extracting structured facts from messy natural language, and that LLMs still hallucinate during that extraction step, meaning the system is only as clean as its input. A few commenters shared their own attempts at similar solutions—one uses a decision log stored in a CLAUDE.md file, another builds linters that enforce output structure, and someone else pointed to Answer Set Programming as a more natural fit for non-monotonic "unlearning" of facts. The thread leaned heavily into the insight that this approach fundamentally separates the probabilistic (natural language understanding) from the deterministic (logical inference), and while nobody claimed it's production-ready, there was broad agreement that this direction—treating LLM memory as a program analysis problem rather than a retrieval problem—is the right way to make agents that can sustain complex, multi-hour investigations without losing track of what's actually true.
Run Qwen3.8 27B locally: real numbers from my Mac Studio [comments]
105 points · 84 comments · terminalbytes.com · 16h ago
A Mac Studio owner published detailed benchmarks of running the new Qwen 3.8 27B model locally, finding it generates at half the tokens-per-second of its predecessor but actually finishes answers faster because it's much more concise, and also documented a bizarre 1-bit "brain damage" quantization that preserves factual recall but makes the model unable to commit to a final answer. The HN thread immediately latched onto that 14 tokens/sec figure as painfully slow, with several people declaring anything under 100 tok/sec worthless for interactive use, though others pushed back hard that 50-60 tok/sec is faster than you can read and that focusing solely on decode speed misses the point for agentic workflows where the model spends most of its time reading tokens. A major split emerged between Mac users frustrated by the speed gap between Qwen 3.8 and 3.6 on identical hardware—blamed on unoptimized Metal kernels for the new hybrid attention architecture—and people running old AMD datacenter cards or RTX 5090s who are getting 70-160 tok/sec for a fraction of the Mac Studio's price. The author of Draw Things chimed in to argue the local LLM scene desperately needs a "just works" Mac app like his image generation tool, sparking pushback that even Draw Things isn't actually beginner-friendly, while others pointed to the constant need to switch between Ollama, LM Studio, llama.cpp, and oMLX just to find the fastest runtime for a given model and quant.
Corporate profits hit highest share since WWII, as worker payouts wilt [comments]
104 points · 18 comments · www.ft.com · 14h ago
The linked article wasn't available to this summarizer; from the discussion, it’s an FT piece showing corporate profits as a share of the economy hitting their highest level since WWII while labor’s share shrinks. The thread mostly takes that as a given and argues over what to do about it, directly grappling with the idea that automation and AI make the old tax base irrelevant. A sharp split emerged: some insist you can’t tax servers or algorithms, while others fire back with concrete mechanisms—taxing electricity consumption per token, seizing shares into a sovereign wealth fund, or just regulating buybacks and dividends more aggressively. Someone also pushed back on the claim that widening stock ownership softens the blow, pointing out the top 10% effectively own the entire market.
Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment [comments]
100 points · 25 comments · arxiv.org · 15h ago
This paper presents a multi-agent AI system called the Station, where different LLM-powered agents collaborate without central coordination to tackle open mathematical problems, claiming novel results on five established problems including new Kakeya sets, kissing configurations, and an improved bound for Erdős's minimum-overlap problem. The HN thread is split between genuine excitement and sharp skepticism: several mathematicians push back hard on the framing, arguing the system is essentially doing rapid recombination of known techniques at scale rather than genuine conceptual discovery, with one person noting "having a vast accessible memory and the ability to very rapidly test/recombine previously-elucidated approaches" is impressive but doesn't settle the open question about whether that counts as creativity. There's a lively side debate about whether anthropomorphizing the system—calling its processes "thinking" or "holidays"—is a harmful distortion or a useful way to erode the specialness people reserve for human cognition. Several commenters point out the agents aren't long-lived due to context limits, which undermines comparisons to human research careers, and the thread also branches into a recommendation for Greg Egan's *Permutation City* and *Diaspora* as relevant sci-fi on automated truth-seeking.
Don't use musl if you care about performance [comments]
94 points · 59 comments · blog.brokk.ai · 16h ago
The article is a performance analysis from a JVM-background developer who switched his Rust projects to musl for the convenience of static binaries, only to discover that musl's allocator is catastrophically slow—144% slower than glibc in his benchmarks—and that swapping in mimalloc only brings the regression down to 26% because other musl routines like memcpy and memset are also slow. The HN crowd largely agrees that musl’s design tradeoffs are well-known, but splits sharply on whether 26% is a disaster or a reasonable price for simplicity and smaller binaries; several ops-adjacent commenters push back that if you have 128GB Kubernetes hosts, just ship glibc containers and stop making the tradeoff. A deep technical tangent erupts around why compilers can't just fix slow memcpy—experts explain that glibc does runtime dispatch to a dozen microarchitecture-tuned assembly implementations that no compiler-builtin can match, while musl prioritizes tiny code size over speed. A few people make the case that if your application allocates heavily enough for the allocator to be the bottleneck, you’re already doing something wrong and should manage your own memory, which other commenters dismiss as overly dismissive of real-world workloads.
Generated 2026-08-29 08:14 UTC
Generated by Sauron from Hacker News discussions and linked articles.