0:00 / 1:23ai Open GLM model beats GPT-5.5 on hallucination rate, but HN isn’t buying the scaling claims
An article argues that Z.ai’s MIT-licensed GLM-5.2, a 753-billion-parameter mixture-of-experts model with about 40 billion active parameters, comes close to larger proprietary models on Artificial Analysis’s Intelligence Index while hallucinating far less on the AA-Omniscience benchmark. It cites hallucination rates of 28% for GLM-5.2, 36% for Opus 4.8, 48% for Claude Fable 5, 86% for GPT-5.5, and 94% for DeepSeek V4 Pro, then argues that raw parameter scaling is hitting diminishing returns and that model selection should weigh capability, uncertainty calibration, and efficiency. HN commenters pushed back hard on the causal leap from these numbers to “bigger is worse,” especially because GPT-5.5 reportedly had higher overall accuracy on the same benchmark while being worse at abstaining when wrong.
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 become worse. The dominant mood was methodological caution: people noted that hallucination rate is conditional, differs from overall accuracy, and may reflect training, eval design, compute, or abstention incentives rather than parameter count alone. (Skepticism of broad claims about scaling limits, Hallucination rate versus overall accuracy, Training and post-training matter more than raw model size)
0:00 / 0:21ai Switzerland pitches Apertus as a fully open sovereign AI model
The Swiss AI Initiative, a collaboration between EPFL, ETH Zurich, and CSCS with Swisscom as a strategic partner, introduced Apertus as an open foundation model aimed at 'sovereign AI.' The project says it will publish weights, training data, code, methods, and alignment principles, with 8B and 70B parameter models trained on more than 1,000 languages and designed around EU AI Act compliance, including opt-outs, PII removal, and anti-memorization measures. The pitch matters because it frames openness not just as developer access, but as a reproducible, legally documented alternative to closed commercial AI stacks.
Discussion: Mixed — HN broadly likes the principle of a fully open, reproducible foundation model and sees strategic value in non-US AI infrastructure, but commenters are skeptical about competitiveness, practical usability, and legal/compliance claims. The discussion repeatedly contrasts Apertus with other open models, debates whether 'sovereign AI' matters beyond ordinary openness, and worries that local/open models still lose to hosted services on user experience. (open weights, open data, and reproducible training are valued, skepticism that Apertus is competitive with leading open models, sovereign AI framed as a response to geopolitical and data-trust concerns)
0:00 / 0:36ai LLMs Have Entered Their Complicated Architecture Era
Ian Barber argues that LLMs, once relatively clean stacks of Transformer blocks, are becoming much more like recommendation systems: full of specialized attention variants, routing, multimodal encoders, MoE components, and distributed inference constraints. The piece says this makes research iteration harder, because new ideas need optimized-enough implementations before teams can even tell whether they are worth pursuing. Barber points to composable kernel work like PyTorch FlexAttention as the kind of tooling needed to keep experimentation practical without giving up performance.
Discussion: Mixed — HN readers were broadly receptive to the article’s core point that modern LLM work is becoming less about simple scaling and more about hard systems engineering, but they debated whether the chosen model comparison overstated the architectural shift. A large side thread got distracted by accusations of AI-written prose and the difficulty of identifying it, while a few commenters complained the article was too technical or the blog was hard to navigate. (LLM architecture complexity is increasing as scaling gains flatten, Performance optimization can become a barrier to architectural experimentation, Debate over whether Nemotron versus Llama is a fair comparison)
0:00 / 0:52biotech A 15-minute at-home test for Lyme-infected ticks
LymeAlert, a $40 at-home tick test due to go on sale in August, is designed to detect Lyme disease bacteria in a crushed tick within 15 minutes using chemically treated paper. The pitch is that a fast result could help bite victims decide whether to seek prompt antibiotic prophylaxis within the recommended 72-hour window, while avoiding unnecessary doctor visits and antibiotics when ticks test negative. Experts quoted in the article say the idea could be useful if accurate, but warn that false positives could cause panic and that the test does not detect other tick-borne hazards such as the agent associated with Alpha-gal syndrome.
Discussion: Mixed — HN’s discussion is less about the product mechanics and more about the anxiety and uncertainty around tick-borne disease. Many commenters share long, personal stories of fatigue, brain fog, sleep problems, and difficult diagnoses, while others caution that similar symptoms can come from long COVID, stress, hormones, sleep disorders, or other conditions. The mood is cautiously interested in faster tick testing, but wary of false positives, incomplete coverage of other pathogens, and the murky territory around chronic Lyme claims and treatments. (Personal stories of suspected or diagnosed Lyme disease, Concern about chronic symptoms and diagnostic uncertainty, Interest in faster treatment decisions after tick bites)
0:00 / 0:46biotech HPV vaccine drives cervical cancer deaths near zero in young women
A Lancet study using England vaccination and mortality data finds that women offered HPV vaccination in early adolescence have seen major reductions in cervical cancer mortality, with no deaths among women aged 20 to 24 in England between 2020 and 2024. Researchers estimate the program has prevented nearly 200 deaths in young women so far, and that those vaccinated at ages 12 or 13 have an almost zero chance of dying from cervical cancer before 30. The warning is that HPV vaccine uptake has fallen since the pandemic, from near the WHO target to about 75% nationally and 60% in London, which could mean avoidable deaths return.
Discussion: Mixed — The discussion is broadly encouraged by the vaccine’s impact, but pushes hard on framing: several commenters argue deaths under 30 were already very rare and want absolute-risk context. Much of the thread shifts to access, eligibility and cost for older adults and men, plus reminders that HPV vaccination also reduces risk for other HPV-related cancers. (Strong support for HPV vaccination as a cancer-prevention success, Concern that the headline overstates impact without absolute baseline risk, Calls to include avoided treatment, not just avoided deaths)
0:00 / 0:52science Why doomscrolling feels like a threat response
A ScienceDaily piece argues that rising news avoidance is not apathy but a predictable collision between human negativity bias and a global, always-on information environment. It cites the Reuters Institute’s 2025 Digital News Report, research showing negative words increase news click-throughs, and a 2022 study in which 17 percent of American adults met criteria for severe Problematic News Consumption. The article says the answer is not total avoidance, but more deliberate consumption: fixed news windows, depth over volume, trusted sources, and connecting information to actions people can actually take.
Discussion: Mixed — HN broadly accepted the premise that constant negative news and social feeds exploit human threat detection, but commenters split on the remedy. Many argued for limiting exposure to events outside one’s control and focusing on local action, while others pushed back that civic awareness, voting, donations, and personal choices still matter. The thread also veered into media literacy, performative activism, policy overreaction, and skepticism about whether avoidance is realistic for people directly affected by the news. (Negativity bias and attention capture, News avoidance versus civic responsibility, Agency, local action, and mental health)
0:00 / 0:42science Slow Breathing May Nudge the Brain Toward Risk
A Neuron paper reports that slow breathing, especially with prolonged exhalation, can modulate brain and autonomic function and shift reward-related behavior, including greater risk-taking. The finding matters because breathing exercises are widely used for anxiety, performance, and emotional regulation, but the behavioral effects may be more nuanced than simply “calming down.” HN readers connected the work to public speaking, tactical breathing, pranayama, and panic management, while questioning how to interpret increased risk tolerance clinically.
Discussion: Mixed — HN was broadly interested but cautious: many commenters found the result plausible through personal experience with public speaking, tactical breathing, yoga, and anxiety control, while others were surprised that parasympathetic activation might increase risk-taking rather than simply produce calmer, more cautious behavior. The thread leaned practical and anecdotal, with debate over mechanisms, propranolol, exposure training, and whether the paper’s framing overstates the benefits. (slow breathing as a tool for public speaking and anxiety, prolonged exhalation, vagal tone, and parasympathetic activation, risk-taking interpreted as reduced fear or increased reward sensitivity)
CSSQuake is a browser-based recreation of Quake that uses CSS for rendering, with the runtime game logic implemented in TypeScript. The project struck a nerve on Hacker News because it turns a famously performance-sensitive 1990s 3D game into a playful demonstration of what modern web layout and styling machinery can be pushed to do. Commenters also surfaced the source repository and noted that the build process extracts facts from QuakeC/progs.dat into JSON for Quake-like gameplay.
Discussion: Positive — HN was broadly delighted and impressed, treating CSSQuake as a joyful technical stunt rather than a practical engine. The main criticism was performance variability, especially reports that Safari/WebKit felt clunky while Firefox and Chrome often ran smoothly. Several commenters clarified that CSS handles the rendering while TypeScript handles the game logic, and the thread drifted into nostalgia for original Quake hardware requirements. (Impressive CSS/PolyCSS showcase, Browser-dependent performance, with Safari/WebKit complaints, Not pure CSS: TypeScript drives game logic)
0:00 / 0:22software Linux finally removes strncpy after six years of cleanup
Linux 7.2 has removed the kernel’s strncpy API after roughly six years and about 362 commits spent eliminating in-kernel users. The function had long been deprecated because its NUL-termination behavior is easy to misuse and because it can waste time zero-filling destination buffers. Kernel code is now expected to use more explicit alternatives such as strscpy, strscpy_pad, strtomem_pad, memcpy_and_pad, or memcpy depending on the destination and padding semantics.
Discussion: Positive — HN is broadly approving, with commenters treating the removal as unglamorous but important engineering work. The discussion is also full of long-running frustration with C strings, NUL termination, and legacy standard-library APIs, plus some debate over whether the replacement APIs are clearer or too fragmented. (Praise for long-term maintenance and bug reduction, Frustration with C string design and NUL-terminated strings, Debate over safer string abstractions such as fat pointers, ranges, and Pascal-style strings)
0:00 / 0:27software A reverse-engineered F-15 Strike Eagle II needs DOS test pilots
A hobby project reconstructing the C source for the 1989 DOS flight sim F-15 Strike Eagle II says it has reached a major milestone: all C code has been reconstructed for all executables, and most assembly-only code now has C replacements. The project is asking owners of version 451.03 with the Desert Storm expansion to test the v0.9.1 DOS build in DOSBox or real DOS, looking for crashes, graphics glitches, broken keys, and other regressions compared with the original. This matters as a preservation and porting step: the team is moving from opcode-matching reconstruction toward maintaining a playable game and eventually forking into a porting project.
Discussion: Positive — The thread is warmly supportive, with lots of nostalgia for MicroProse-era flight sims and respect for the reverse-engineering effort. Discussion centers on why source reconstruction matters when DOSBox already exists, plus practical notes on portability, legality, and whether AI helps with decompilation. (nostalgia for DOS flight simulators, source reconstruction as preservation and modding, future ports beyond DOS)
0:00 / 0:28software TownSquare makes websites feel inhabited — and immediately hits the moderation wall
TownSquare is a small embeddable presence layer that lets visitors to a website appear together, move around, chat briefly, and interact with props via a single snippet. The pitch is a more inhabited, social web without accounts or algorithms. On HN, the concept landed as charming and fun, but the public demo was quickly overrun by spam and offensive behavior, making moderation the central issue.
Discussion: Mixed — HN broadly liked the playful, nostalgic idea of adding lightweight live presence to ordinary websites, but the live demo quickly became dominated by spam and offensive or juvenile messages. The discussion shifted from appreciation of the design to practical concerns about moderation, abuse, bots, rate limits, and whether site owners would risk embedding an unmoderated public layer. (Playful web presence and real-time visitor interaction, Unmoderated anonymous chat becoming spammy under attention, Need for site-owner controls, filters, bans, and rate limiting)
0:00 / 0:39software Linux async I/O: epoll’s readiness model versus io_uring’s completion model
The linked writeup compares Linux’s long-standing epoll interface with io_uring, using a student-built reverse proxy project as the motivation. It argues that epoll’s readiness notifications still require separate read and write syscalls, while io_uring uses shared ring buffers to submit batches of operations and receive completions, potentially reducing syscall overhead and enabling features like registered buffers, zero-copy sends, and SQPOLL. The practical importance is that switching from epoll to io_uring is not just an API swap; it changes the architecture and error-handling model, and may require a rewrite to see the promised performance gains.
Discussion: Mixed — HN liked the article as an accessible tour of Linux async I/O, but the discussion quickly turned from enthusiasm to caveats: performance depends heavily on workload, CPU affinity, batching, zero-copy paths, and framework integration. Several commenters also stressed that io_uring’s adoption is constrained by security history and deployment defaults, so it is not a universal replacement for epoll yet. (io_uring can reduce syscall overhead through batching and completion-based I/O, Real proxy performance also depends on CPU pinning, NIC queue alignment, false sharing, allocators, and kernel networking features, Existing async frameworks may not benefit unless redesigned around io_uring’s completion model, chaining, and SQPOLL-style operation)
0:00 / 1:34security Loupe Shows What iPhone Apps Can See Without Asking
Mysk released Loupe, a free open-source iOS and iPadOS app that displays real values available through public Apple APIs to demonstrate the device fingerprinting surface exposed to native apps. It groups signals into passive data available without prompts, permission-gated data, and more advanced side-channel techniques such as URL-scheme probing and Keychain persistence across reinstalls. The project’s point is that trackers do not need obvious identifiers like a name, email, or location when many small device readings can be combined to recognize a user again.
Discussion: Negative — HN’s reaction is less about Loupe itself and more frustration that iOS exposes enough passive signals for app fingerprinting while giving users no simple way to block network exfiltration. Commenters largely agree the tool is useful and eye-opening, but the mood is critical of Apple’s permission model and the broader app economy. (Calls for per-app internet access permissions or a Little Snitch-style firewall on iOS, Concern that passive public APIs can be combined into a durable device fingerprint, Comparisons with GrapheneOS, LineageOS, AOSP, and mainland-China iPhones that offer stronger network controls)
0:00 / 1:08hardware Proxmox Gets a MicroVM Shortcut
A long-time Proxmox user built pve-microvm, a Debian package that makes QEMU’s microvm machine type feel like a managed Proxmox guest. It patches Proxmox’s qemu-server internals, adds web UI integration, ships a small custom Linux kernel and initrd, and can boot minimal isolated guests in roughly sub-300 milliseconds, with support for a wide range of guest OSes. The pitch is a middle ground between fast but weaker-isolated LXC containers and heavier full VMs: KVM isolation without BIOS, GRUB, VGA, legacy chipset emulation, or slow device probing.
Discussion: Positive — HN is broadly impressed by the technical work and excited about first-class microVMs in Proxmox, especially for homelabs, coding-agent sandboxes, and lightweight isolated services. The main hesitation is operational risk: pve-microvm patches Proxmox internals, so commenters worry it could break across updates and would prefer upstream support. Discussion also branches into alternatives like Firecracker, libkrun, Kata Containers, microvm.nix, SlicerVM, and the still-unsolved pain around GPU and CUDA support in microVMs. (Strong interest in VM-grade isolation with container-like boot times, Desire for official Proxmox support rather than Perl-module patching, Comparisons with Firecracker, libkrun, Kata Containers, KubeVirt, and microvm.nix)
0:00 / 0:42startups Was This Startup Built to Ship—or to Siphon Fees?
A former GenieDB engineer revisits the startup that brought him to the U.S. after learning that the VC fund behind it, Frost VP, later faced SEC action tied to alleged excessive incubator fees. Digging through arbitration and SEC records, he concludes that while the team was genuinely trying to build database technology, the fund may have been using portfolio companies like GenieDB to extract investor money through fees. The piece matters because it frames startup failure not just as bad strategy or bad luck, but as something that can reshape employees’ lives while barely registering in the legal record.
Discussion: Negative — The HN discussion is mostly cynical and resigned, using the GenieDB story as a jumping-off point for broader complaints about perverse incentives, outsourcing markups, budget games, and fraud-adjacent corporate behavior. Several commenters share firsthand stories where jobs, projects, or entire companies seemed to exist less to build value than to route money through accounting or contracting structures. (perverse incentives in startups and large organizations, outsourcing and contractor markups, budget-use-it-or-lose-it behavior)
0:00 / 0:42startups StartupWiki pitches a free Crunchbase alternative, but HN wants proof
A maker launched StartupWiki, a free startup database meant to feel more like Wikipedia than Crunchbase: no subscriptions, no accounts required for browsing, and startup profiles with search, filtering, categories, and a public API in progress. The project is early, and HN’s response was that the idea is useful only if the data becomes broad, current, licensed, forkable, and verifiably sourced. Several users found missing companies and incorrect or outdated funding details, while the founder responded in-thread with fixes, plans for better agent citations, and manual approvals for new listings.
Discussion: Mixed — HN liked the public, no-paywall startup database idea, but the discussion quickly centered on trust: missing companies, wrong funding data, unclear verification, licensing, and whether AI-generated profiles can be reliable. Commenters offered concrete ideas around open sourcing, data dumps, source-linked verification, Wikidata integration, machine-readable startup metadata, and better login/API flows. (Strong demand for provenance and source-linked verification, Concern that AI-generated startup profiles are inaccurate or stale, Requests for open source code, clear content licensing, and downloadable data dumps)
0:00 / 1:31policy Claude’s ID checks spark privacy backlash
Anthropic’s Claude support site says the company is rolling out identity verification for certain capabilities, 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, not train Anthropic models, but commenters focused on Persona’s role, possible fraud-prevention data use, and legal-process disclosures. The story matters because access to frontier AI models is increasingly being tied to real-world identity, raising questions about privacy, export controls, and who gets access to the most capable systems.
Discussion: Negative — HN reaction is strongly negative, with many commenters saying they canceled or may cancel Claude over government-ID and selfie verification through Persona. The main pushback is privacy and surveillance risk, especially around third-party data handling, legal requests, and possible ties to export-control or model-access restrictions. A smaller corrective thread notes that this support page and ID process appear to have existed for months, so commenters may be over-linking it to the latest Fable controversy. (Privacy concerns over handing IDs and selfies to Persona, Fear of government surveillance or export-control-driven access limits, Frustration with unclear triggers for verification and account bans)
0:00 / 1:51policy A slick AI fan site copied an entire bestselling book
Andy Baio reports that an unauthorized site at a lookalike domain reproduced John Koenig’s The Dictionary of Obscure Sorrows, including the book’s foreword and all 311 entries, while replacing the original illustrations with DALL-E 2 images and adding a GPT-4-powered “generate your own words” feature. Koenig told Baio he had nothing to do with the site, which credits San Francisco web and marketing agency Qontour and appears to have been used as a portfolio showcase. The case matters because it blends old-fashioned wholesale copying with AI-era presentation, raising questions about how creators can realistically enforce rights when infringement can be packaged quickly and professionally.
Discussion: Negative — HN reaction is strongly critical of the alleged copying and deeply pessimistic about copyright enforcement for independent creators. Commenters focus less on the AI imagery itself and more on the site reportedly reproducing the book verbatim, confusion around DMCA processes, and the sense that platforms respond unevenly depending on who is making the claim. (AI tools lowering the cost of copying and rebranding creative work, DMCA takedowns seen as ineffective or unevenly enforced, Copyright enforcement favoring large rights-holders over individuals)
0:00 / 1:16general A Free RTS Revives Total Annihilation’s Giant Robot Wars
Beyond All Reason is a free real-time strategy game inspired by Total Annihilation, emphasizing fully simulated units, projectiles, explosions, ballistics, terrain deformation, and battles involving huge armies. The project is drawing attention on HN both as a technically impressive RTS revival and as a rare modern entry in a genre many commenters feel has faded. The main caveat from the discussion is that the popular online team modes can be unforgiving, with multiple players describing a steep learning curve and sometimes toxic lobby culture.
Discussion: Mixed — HN is enthusiastic about Beyond All Reason’s technical ambition, scale, and Total Annihilation lineage, but the dominant discussion is a warning about multiplayer culture. Many commenters praise the game as mature, performant, and unusually strong for a free RTS, while others say the standard 8v8 scene can be hostile to new or casual players. Several users recommend solo, PvE, smaller matches, noob-friendly lobbies, or map-rotating lobbies as better entry points. (Nostalgia for Total Annihilation and 1990s RTS games, Technically impressive large-scale simulation, Toxicity and rigid meta expectations in competitive team lobbies)
0:00 / 0:25general Google says IPv6 has crossed the halfway mark
Google’s measurements show IPv6 reaching 50% of users accessing Google services for the first time, while APNIC’s separately weighted global measurement puts worldwide IPv6 capability at 42%. APNIC says the gap comes from methodology: Google measures its own service access, while APNIC uses ad-based tests weighted by estimated Internet population per economy. The milestone matters because IPv6 is now mainstream at global scale, but adoption remains highly uneven across countries, ISPs, devices, and services.
Discussion: Mixed — HN treats the 50% mark as a real milestone, but the mood is more weary than celebratory. Commenters focus on holdout ISPs, IPv4-only major services, awkward home-router behavior, and the uneven incentives that keep dual-stack networking around. (50% is meaningful but measurement-dependent, ISP holdouts remain a major frustration, Client-side IPv6 adoption is ahead of server-side adoption)
0:00 / 0:22general Finland’s libraries lend sewing machines — and a model of civic life
The BBC profiles Finland’s expanding view of public libraries, centered on Helsinki’s Oodi Library and similar services in Oulu, where residents can borrow or book far more than books: sewing machines, rooms, 3D printers, game collections, music spaces, sports gear, and even help navigating online bureaucracy. The article argues that Finland treats libraries as democratic infrastructure, backed by law and relatively high public spending: nearly €371 million in 2025, or €65.78 per person. For HN readers, the story resonated less as a Finnish oddity than as a familiar and desirable model: many commenters cited local libraries already lending tools, instruments, telescopes, kayaks, park passes, and maker-space equipment.
Discussion: Positive — HN was broadly enthusiastic about libraries as shared civic infrastructure, with many commenters listing local “library of things” examples in the US and Canada. The main cautions were operational: high-demand items can have long waits, complex tools need maintenance and staff expertise, and some libraries struggle with being de facto shelters when broader social services fail. (Libraries of Things and tool lending, Libraries as third spaces and democratic infrastructure, Maker spaces, 3D printers, sewing machines, instruments, and passes)
0:00 / 0:41general HN Marks Father’s Day With Maker-Dad Memories
A Tell HN post for Father’s Day turned into a community memory thread, sparked by the poster’s story of an uncle who filled a father role in Soviet-controlled Poland and built hackerish childhood experiences, including rocket cars and a working police siren on a banana bike. The discussion matters less as news than as a snapshot of HN’s culture: technical curiosity, DIY problem-solving, and mentorship framed as family inheritance.
Discussion: Positive — The thread is overwhelmingly warm and reflective, with many HN users sharing first Father’s Day stories, memories of technically curious dads, and gratitude for family role models. A few comments add bittersweet notes about loss, infertility, early-parenthood struggle, and the fact that Father’s Day is not observed on the same date everywhere. (gratitude for fathers and father figures, parenthood as transformative and rewarding, maker culture passed between generations)