0:00 / 1:19ai Open GLM model beats GPT-5.5 on hallucination rate, but HN isn’t buying the scaling claim
The post argues that Z.ai’s MIT-licensed GLM-5.2, with 753 billion parameters and roughly 40 billion active, comes close to proprietary frontier models on Artificial Analysis’s intelligence index while hallucinating far less on AA-Omniscience: 28% versus 86% for GPT-5.5, according to the article. It uses that gap, plus a coding example where GLM-5.2 recognized an impossible prompt, to argue that raw model size and test-time reasoning budgets are no longer reliable proxies for truthfulness. The discussion matters because it gets at a central AI product question: whether users should prefer maximum capability, better uncertainty calibration, or cheaper, more reliable behavior.
Discussion: Mixed — Commenters were interested in the reported hallucination gap, but largely skeptical of the article’s broader conclusion that bigger models have plateaued or get worse. The dominant mood was methodological caution: people noted that hallucination rate, overall accuracy, abstention behavior, benchmark design, post-training, and real-world task distribution are different things. (Skepticism about claims that model scale has plateaued, Hallucination rate versus overall accuracy, Training and reinforcement learning incentives against saying “I don’t know”)
0:00 / 0:21ai Switzerland’s Apertus pushes fully open sovereign AI
The Swiss AI Initiative—backed by EPFL, ETH Zurich, CSCS, and strategic partner Swisscom—has introduced Apertus, a foundation model effort positioned around “open weights, open data, open science.” The project says its 8B and 70B-parameter models are multilingual across 1,000-plus languages and designed for EU AI Act compliance, including opt-outs, PII removal, and anti-memorization measures. The significance is less just another LLM release and more Europe’s attempt to build reproducible, legally conscious AI infrastructure outside the dominant US and Chinese ecosystems.
Discussion: Mixed — HN is broadly supportive of the fully open approach and the strategic value of non-US AI capability, but skeptical about Apertus’s current model quality and pace. Commenters repeatedly compare it with OLMo, Nemotron, Chinese open models, and prior Apertus releases, with several saying openness may matter more than benchmark leadership for science and sovereignty. (Praise for open weights, data, code, training recipes, and reproducibility, Skepticism about competitiveness and reports of hallucinations or weak multilingual behavior, Sovereign AI framed as a response to US and Chinese platform dependence)
0:00 / 0:22ai Open models are close enough to quit Claude, says one developer
Andrew Marble argues that switching from Claude and GPT-style proprietary models to open-weight models now has a much smaller productivity penalty than it used to, comparing the moment to Linux becoming practical for mainstream work. His trigger is Claude’s identity verification rollout, plus broader concerns about safeguards and data retention; he says open models are close to the leaders, usable through coding harnesses, and can be run locally or in the cloud. The piece matters because it captures a growing developer calculation: whether model quality gains from the top closed systems still outweigh privacy, control, and platform-dependence costs.
Discussion: Mixed — HN was broadly sympathetic to the idea of reducing dependence on Anthropic and OpenAI, especially after Claude identity verification and retention concerns, but commenters split hard on whether open-weight models are actually good enough for serious coding work. The dominant mood was pragmatic: open models offer control, routing flexibility, and privacy options, but still carry real tradeoffs in performance, cost, hosting complexity, and trust in inference providers. (open-weight models may be only months behind proprietary leaders, privacy and data residency drive interest in self-hosting and EU inference providers, skepticism that benchmarks match real-world coding performance)
0:00 / 0:40ai Open-weights GLM 5.2 takes on Claude Opus in a WebGL coding test
TechStackUps compared Z.ai’s new open-weights GLM-5.2 against Claude Opus 4.8 on a single long-running task: build a raw WebGL 3D platformer from scratch using the same prompt and assets. In the author’s run, Opus finished in about 33 minutes versus GLM’s roughly 71 minutes and produced a cleaner game, while GLM cost $5.39 billed versus an estimated $21.92 for Opus. The takeaway was not that GLM replaces Opus, but that a capable MIT-licensed open model with a 1M-token context window is now good enough to keep in the coding-agent toolkit.
Discussion: Mixed — HN was impressed that GLM-5.2 could produce substantial coding output at much lower cost, especially without multimodal vision, but the dominant mood was skeptical about the methodology. Commenters repeatedly argued that one-shot greenfield demos are useful vibe checks, not benchmarks, and wanted tests of collaboration, steerability, reliability, and work inside existing codebases. (Skepticism toward one-shot prompting as a benchmark, Interest in open-weights models as cheaper, durable alternatives, Opus seen as faster, more steerable, and better for collaborative coding)
0:00 / 0:52biotech A fifteen-minute home test for Lyme-carrying ticks
LymeAlert, a $40 at-home tick test due to go on sale in August, is designed to tell users in about 15 minutes whether a removed tick carries Lyme disease bacteria. The test grinds up as many as five ticks and uses treated paper that changes color if the bacteria are present. Its promise is faster decisions about preventive antibiotics within the recommended 72-hour window, but experts quoted in the article warn that false positives could cause panic and that the test does not detect other tick-borne hazards.
Discussion: Mixed — HN’s discussion is less about the product’s mechanics and more about the fear and ambiguity around tick-borne illness. Many commenters shared long, unresolved health stories involving Lyme, long COVID, sleep, inflammation, hormones, or mold, while others urged caution about internet diagnoses and unproven treatments. The mood is hopeful that faster tick testing could help, but wary of false positives, missed co-infections, and the broader difficulty of diagnosing chronic symptoms. (Personal accounts of suspected or diagnosed Lyme disease, Concern over chronic fatigue, brain fog, anxiety, and sleep problems, Skepticism toward fringe or expensive treatments)
0:00 / 0:35biotech FDA advisors back Moderna’s mRNA flu shot after a blocked review
FDA vaccine advisors voted 9–0 to recommend approval of Moderna’s seasonal mRNA flu vaccine, mRNA-1010, branded mFlusiva. Trial data presented to the panel showed it was about 27 percent more effective than a standard flu shot in adults 50 and older, and produced stronger immune responses than a high-dose flu vaccine in a smaller study of adults 65 and older. The vote follows an earlier controversy in which then-FDA vaccine official Vinay Prasad refused to review Moderna’s filing, a decision later reversed by the agency. FDA’s final decision is due by August 5, and CDC recommendation and coverage decisions remain a further hurdle.
Discussion: Mixed — Commenters were broadly supportive of the advisory panel’s unanimous vote and of mRNA vaccine technology, but the discussion was dominated by anger and distrust over political interference in FDA decision-making. Many framed the earlier refusal to review the vaccine as anti-science or ideologically driven, with some concern that approval and CDC recommendations could still be derailed. (Frustration with political interference in science agencies, Support for evidence-based vaccine review and mRNA platforms, Concern about public distrust of vaccines after COVID-era controversies)
0:00 / 0:43science Your brain was not built for the doomscroll
A ScienceDaily piece argues that rising news avoidance is not apathy but a predictable psychological response to a threat-sensitive brain being asked to process global crises in real time. It cites Reuters Institute data showing 69 percent of Canadians and 40 percent globally at least sometimes avoid news, along with research linking negative headlines to higher click-through rates and severe “problematic news consumption” to worse well-being. The proposed fix is not total avoidance, but bounded, higher-quality news consumption paired with concrete action where possible.
Discussion: Mixed — HN broadly accepts the idea that nonstop negative news exploits human threat-detection, but the discussion quickly turns into a debate about agency: when to unplug, when to stay informed, and whether news-driven anxiety reflects media design or unrealistic expectations about risk and politics. (Negativity bias and attention capture, Doomscrolling and social media incentives, Limits of individual agency over global crises)
0:00 / 0:20science Slow Breathing May Nudge the Brain Toward Taking More Risks
A Neuron paper reports that slow breathing, particularly prolonged exhalation, can modulate brain function and shift behavior around reward and risk. The finding matters because breathing exercises are widely used for anxiety and performance, but the study suggests their effects may include changing reward responsiveness and risk tolerance, not simply making people calmer.
Discussion: Mixed — HN was broadly intrigued and personally engaged, with many commenters sharing breathing techniques for public speaking, anxiety, athletics, and stress. The main surprise was that parasympathetic activation and prolonged exhalation were linked not just to calm, but to increased reward responsiveness or risk tolerance, which some saw as useful and others found counterintuitive or potentially over-sold. (Slow breathing as a practical anxiety and public-speaking tool, Debate over whether increased risk-taking is beneficial or concerning, Connections to vagal tone, parasympathetic activation, and reward processing)
0:00 / 0:36science A Math Essay Reframes Logarithms as Units, Vectors, and Torsors
The linked essay proposes treating a logarithm without a base as an abstract quantity, with ordinary base-specific logs emerging as ratios—much like measuring the same geometric quantity in different units such as bits, nats, or digits. It then extends the analogy toward vectors, coordinate systems, and other mathematical structures, arguing that many appearances of logarithms share a deeper pattern. The discussion matters less as a new theorem than as a pedagogical and conceptual attempt to make change-of-base, units, and abstraction feel more natural.
Discussion: Mixed — HN was engaged and mostly appreciative of the mathematical reframing, especially where commenters connected the essay’s “baseless logarithm” idea to torsors, units, and older computational uses of log tables. But there was also skepticism that the piece was more notational play than new math, plus criticism that later sections needed stricter type or dimensional discipline. (Baseless logarithms as torsors or unit choices, Logarithms as practical computational tools, including log tables and slide rules, Debate over whether the essay is insight or notation play)
0:00 / 1:20software Deno Takes a Swing at Electron
Deno has previewed “Deno Desktop,” a coming Deno 2.9 feature that packages Deno projects — from a single TypeScript file to frameworks like Next.js, Astro, Fresh, Remix, Nuxt, SvelteKit, SolidStart, TanStack Start, or Vite SSR — as self-contained desktop apps. It bundles app code, the Deno runtime, and a rendering backend, with a small-by-default system WebView path and an optional bundled Chromium/CEF backend for consistent rendering. The feature is only in the canary build for now, and its CLI, config keys, and TypeScript APIs may still change before stable release.
Discussion: Mixed — The HN mood is broadly interested and cautiously positive, especially around having another Electron/Tauri alternative with Deno’s tooling, Node compatibility, framework detection, and optional CEF. But the thread quickly turns into familiar desktop-web tradeoff debates: bundled Chromium versus system webviews, binary size, native look-and-feel, permission semantics, and whether the claimed in-process binding model is really avoiding IPC. (Interest in Electron and Tauri alternatives, Debate over system webviews versus bundled Chromium/CEF, Questions about binary size and shared CEF runtimes)
0:00 / 0:22software A Photo Hoarder Finds Years of Accidental Wigglegrams
The author discovered that their habit of taking many near-identical photos had accidentally created a large archive of potential wigglegrams: looping stereo-like GIFs made from slightly different viewpoints. They wrote a script using perceptual hashes and Hamming-distance thresholds to find similar image runs in an iCloud or local photo library, then stitch them into animations. The result is part visual experiment, part practical photo-mining tool, and a reminder that modern camera rolls contain lots of latent media beyond the single selected shot.
Discussion: Mixed — HN mostly enjoyed the playful hack and the visual design, but a sizable thread found the looping motion uncomfortable or less convincing than purpose-built stereo images. The technical discussion quickly moved into perceptual hashing, duplicate detection, Live Photos, old stereo cameras, and ways to make the effect smoother or interactive. (Appreciation for a fun, personal image-processing hack, Motion sickness, migraine triggers, and requests for pause or slower animation, Perceptual hashing for finding similar photos and deduplication)
0:00 / 0:23software Linux finally removes strncpy from the kernel
Linux 7.2 has eliminated the kernel’s strncpy API after roughly six years and about 362 commits spent removing remaining users. The function had long been deprecated because its semantics around NUL termination are counter-intuitive, it can cause bugs, and it can waste work by zero-filling destinations. Kernel code is now expected to use more specific alternatives such as strscpy(), strscpy_pad(), strtomem_pad(), memcpy_and_pad(), or memcpy(), depending on whether the destination is a C string, a padded field, or a known-length memory copy.
Discussion: Mixed — Commenters broadly welcomed the cleanup and respected the long, unglamorous engineering effort, but the thread quickly turned into a critique of C string handling, NUL-terminated strings, and standard-library stagnation. The mood was appreciative of the kernel work while frustrated that decades-old API design still creates this much risk and toil. (strncpy as a recurring source of C bugs, praise for long-term maintenance work in critical infrastructure, debate over NUL-terminated versus length-prefixed strings)
0:00 / 0:22software Codex Logging Bug May Hammer SSDs With Terabytes of Writes
A GitHub issue reports that Codex may be continuously writing large volumes of data to its local SQLite feedback log files under `~/.codex`, with one machine showing about 37 terabytes written over 21 days of uptime. The report attributes most retained bytes to global TRACE logging, mirrored OpenTelemetry events, and raw websocket/SSE-related logging, and argues that insert-and-prune behavior in SQLite is amplifying writes far beyond the retained database size. The concern is practical: at the reported rate, a consumer 1 TB SSD could approach its warranted write endurance in under a year if left running heavily.
Discussion: Negative — HN reaction is sharply negative, using this bug as evidence that AI coding tools are shipping with poor engineering polish and resource management. Some commenters note Codex is open source and patchable, and share workarounds, but the dominant mood is frustration with Codex, Claude Code, Electron-style desktop apps, and the broader quality of agentic coding products. (Heavy disk writes and SSD endurance concerns, Complaints about Codex Desktop performance, GPU, memory, and token use, Comparisons with Claude Code and other AI coding harnesses)
0:00 / 0:19software TownSquare makes websites feel inhabited — and immediately tests moderation
TownSquare is a small embeddable presence layer for websites: add a snippet, and visitors appear together on the page, move around, send short messages, and interact without accounts. The concept resonated because it brings back a playful, social feeling to otherwise static websites, but the HN-driven demo was quickly overrun by spam and offensive messages. That made the launch less a pure product showcase and more a live demonstration of why even tiny social features need moderation from day one.
Discussion: Mixed — HN liked the playful, lightweight idea of adding real-time presence to ordinary websites, but the live demo quickly became a case study in anonymous abuse, spam, and moderation challenges. Many commenters said they would only use it with site-owner controls, filtering, bans, or some form of client-side or AI-assisted moderation. (Playful ambient presence on the web, Anonymous public chat quickly attracting spam and offensive behavior, Need for moderation, filtering, bans, and site-owner control)
0:00 / 0:39software JSON-LD for personal sites meets a skeptical SEO crowd
The article explains how personal websites can use JSON-LD with Schema.org to describe entities like a WebSite, WebPage, Person, breadcrumbs, posts, and other page metadata for crawlers. It frames JSON-LD as a way to improve machine understanding, rich previews, and potentially search presentation, while noting practical details like stable @id values and avoiding overly verbose repeated data. The HN discussion turned less on the syntax and more on whether helping crawlers still helps independent publishers in a search world increasingly dominated by AI summaries.
Discussion: Mixed — HN appreciated the practical explanation, but the dominant mood was skeptical about structured data as an SEO strategy in an era where Google and LLM-driven search may summarize content without sending traffic back. Commenters split between pragmatic advice—use Schema.org and follow Google/Bing documentation for specific rich-result use cases—and broader frustration that metadata increasingly benefits platforms more than site owners. (Structured data can help rich snippets, sitelinks, maps, FAQs, and email features when used for supported cases., Many commenters see JSON-LD for SEO as limited, overhyped, or useful only when aligned with search-engine documentation., There is strong resentment toward Google AI overviews and LLM crawlers consuming site content while reducing referral traffic.)
0:00 / 1:34security Loupe shows what iPhone apps can quietly fingerprint
Mysk released Loupe, a free open-source iOS and iPadOS app that displays real values exposed by public Apple APIs and explains how they can contribute to device fingerprinting. It groups signals into passive access, permission-gated data, and advanced side channels such as URL-scheme probing and Keychain persistence across reinstalls. The project is meant as a hands-on privacy awareness tool: Loupe says the values stay on-device unless the user explicitly exports them.
Discussion: Negative — HN reaction is privacy-alarmed and critical of Apple’s defaults. The discussion largely pivots from Loupe itself to why iOS does not offer per-app internet permission or stronger network visibility, with comparisons to GrapheneOS, LineageOS, macOS sandboxing, and iPhones sold in China. (Demand for per-app internet access controls on iOS, Concern that public APIs enable device fingerprinting without prompts, Skepticism that permission prompts alone solve tracking)
0:00 / 1:19startups Was a Startup Job Just a Side Effect of VC Fraud?
A former GenieDB engineer revisits the startup that brought him from the UK to the US after learning that its VC owner, Stuart Frost’s Frost VP, was later accused by investors and the SEC of fraud involving excessive incubator fees. The author reviews arbitration and SEC records and concludes that GenieDB appears to have been used, at least in part, to generate fees for the fund’s incubator operation, even though the engineering team was genuinely trying to build real technology. The piece matters because it puts a personal human frame around a familiar startup-finance question: when a company has little revenue and unclear prospects, who is it really serving?
Discussion: Negative — HN’s reaction is cynical and heavily experiential: many commenters use the story as a springboard for tales of incubator grift, government contracting waste, outsourcing arbitrage, budget games, and companies that seem designed for fees, tax treatment, lawsuits, or optics rather than real products. A smaller thread separates fraud from ordinary startup failure, arguing that failed product-market fit and abandoned products are common and not inherently corrupt. (incubator and VC fee extraction, government and public-private funding waste, contractor and outsourcing budget arbitrage)
0:00 / 0:39startups StartupWiki pitches a free Crunchbase alternative, but HN wants proof and better data
A maker launched StartupWiki, an early free startup database meant to feel more like Wikipedia than Crunchbase: no accounts required, simple company pages, search and filtering, categorization, and a public API in progress. The Hacker News discussion quickly focused less on the concept and more on execution: commenters reported missing startups, incorrect funding and valuation data, confusing “verified” badges, and no visible content license. The founder responded throughout, saying YC data had been scraped, AI-agent profile generation was being improved, source-link bugs were being fixed, and an “agent ledger” had been added.
Discussion: Mixed — Commenters liked the goal of a free, low-friction startup database, but the dominant concern was trust: many found missing companies, outdated or wrong facts, unclear verification, and no visible content license. The founder was active in the thread, acknowledged bugs and pipeline issues, and shipped or promised fixes, which softened the response but did not resolve skepticism about AI-generated data quality. (Strong demand for an open, free startup database, Coverage is currently thin, including missing YC and prominent startups, Verification badges need source links, provenance, and timestamps)
0:00 / 1:11policy Claude ID checks spark privacy backlash
Anthropic’s Claude support page says the company is rolling out identity verification for some use cases, platform integrity checks, and safety or compliance measures, using Persona Identities as its verification partner. The page says verification data is used to confirm identity and not to train Anthropic’s models, but commenters focused on the privacy tradeoff, third-party data handling, and whether ID checks will restrict access to frontier models. The discussion also surfaced an important caveat: multiple users say this help page has been live since April, so the current outrage may be tied to broader anxieties about AI access controls rather than a brand-new announcement.
Discussion: Negative — The HN thread is overwhelmingly distrustful of Anthropic’s identity verification, especially the use of Persona and the possibility that ID checks could become a gatekeeper for advanced AI access. A smaller but important counter-thread notes that the support page and ID process appear to have existed for months, so some reactions may be conflating this with newer model-access controversies. (Privacy and biometric-data risk, Distrust of Persona as verification vendor, Fear of government or export-control-driven AI access tiers)
0:00 / 0:43policy Danish privacy activist says police raided him after posting the prime minister’s numbers
Danish privacy activist and former police officer Lars Kragh Andersen says armed, masked police broke down his door after he obliquely published Prime Minister Mette Frederiksen’s social security number and phone number, and posted a WhatsApp attempt to question her about encryption bans and expanded surveillance. He claims officers cut power to his router, removed Google Nest cameras with local storage, and did not initially tell him the charges. The story matters because it sits at the intersection of anti-surveillance activism, doxxing, law-enforcement tactics, and Denmark’s broader debate over encryption and state access to personal data.
Discussion: Mixed — HN commenters are alarmed by the alleged police tactics—masked entry, cutting power, and seizing cameras—but many are also uneasy about Lars Kragh Andersen’s own methods. The thread splits between people seeing him as exposing government hypocrisy on privacy and surveillance, and others arguing that alleged doxxing, tracking ministers’ cars, or involving families crosses a clear ethical and legal line. (concern over anti-encryption and surveillance policy in Denmark and the EU, debate over whether provocative activism helps or discredits a cause, allegations of selective prosecution and unequal treatment under the law)
0:00 / 1:21general Free RTS Beyond All Reason Revives Total Annihilation at Massive Scale
Beyond All Reason is a free real-time strategy game inspired by Total Annihilation, emphasizing fully simulated units, projectiles, explosions, terrain deformation, and battles involving armies of thousands. The project drew attention on HN as a rare modern RTS with serious scale and polish, especially for fans of the TA and Spring RTS lineage. The comment thread, however, focused as much on the social experience as the game itself, with many users praising the design while warning that public multiplayer can be harsh for new or casual players.
Discussion: Mixed — HN commenters were highly impressed by Beyond All Reason as a technically ambitious, free RTS in the Total Annihilation lineage, with praise for scale, controls, performance, and nostalgia. But the dominant discussion quickly turned to multiplayer culture: several players described the community as toxic, meta-driven, and unfriendly to newcomers, while others argued the problem is mostly confined to popular 8v8 lobbies and can be avoided with solo, co-op, smaller games, or beginner-friendly rooms. (Strong nostalgia for Total Annihilation and 1990s RTS games, Praise for real-time simulation, scale, controls, and performance, Concerns about toxic multiplayer behavior and rigid meta expectations)
0:00 / 0:26general Google Says IPv6 Has Finally Hit Half the Internet
Google’s measurements show IPv6 crossing 50% of user access to Google services for the first time, a symbolic milestone for a protocol that has been rolling out for decades. APNIC cautions that its own global measurement is lower, at 42%, largely because it weights country-level samples by estimated Internet user populations rather than aggregating ad-delivered samples directly. The bigger point: IPv6 is now deployed at global scale, but adoption remains highly uneven by region, ISP, access network, and service provider.
Discussion: Mixed — HN treats the 50% mark as a real milestone, but the mood is more weary than celebratory. Commenters focus on lagging ISPs, IPv4-only major services, home-router pain, and the awkward incentives that keep dual-stack networks around, while also noting that IPv6 can already be faster and cleaner where CGNAT is the IPv4 alternative. (Frustration with ISPs that still do not offer residential IPv6, Measurement nuance: Google shows 50% while APNIC shows 42% due to different weighting methods, Major services and developer infrastructure still lacking IPv6 support)
0:00 / 0:43general Hacker News marks Father’s Day with stories of dads, mentors, and making things
A “Tell HN” post wished the community a happy Father’s Day through a personal memory: the poster’s uncle, acting as a father figure in Soviet-controlled Poland, fostered a hacker-like childhood with rocket cars and an electrically functional police siren mounted to a banana-seat bike. The post resonated because it framed fatherhood broadly—as mentorship, curiosity, mischief, and the adults who help turn kids into makers.
Discussion: Positive — The thread is overwhelmingly warm and reflective, with many commenters sharing first Father’s Day stories, memories of technically curious dads, and gratitude for parents or father figures. The mood is not purely celebratory: several comments acknowledge difficult childhoods, infertility, grief, and the fact that Father’s Day is not universal or easy for everyone. (gratitude for fathers and father figures, new parenthood and early-childhood intensity, technical curiosity passed down through family)