0:00 / 1:16ai Open GLM model beats GPT-5.5 on hallucination benchmark, but HN isn’t buying the scaling takeaway
The article argues that Z.ai’s MIT-licensed GLM-5.2, a 753B-parameter mixture-of-experts model with about 40B active parameters, comes close to proprietary frontier models on Artificial Analysis’s Intelligence Index while showing a much lower hallucination rate on AA-Omniscience. It cites GLM-5.2 at 28% hallucination versus GPT-5.5 at 86%, while also noting GPT-5.5 and other closed models are estimated to be much larger. The author’s broader claim is that the industry should stop treating bigger models and higher benchmark scores as sufficient, and instead optimize for capability, uncertainty calibration, and compute efficiency.
Discussion: Mixed — The discussion is skeptical of the article’s headline conclusion, especially the claim that larger models have plateaued or inherently hallucinate more. Commenters found the GLM-5.2 numbers interesting, but repeatedly warned that hallucination-rate benchmarks are conditional, task-dependent, and not the same as overall accuracy. (Pushback on broad claims about model scaling and parameter count, Debate over hallucination metrics versus overall answer accuracy, Training data quality, RL/post-training, and incentives to guess rather than abstain)
Ian Barber argues that LLMs have moved away from the relatively clean Transformer stacks of early Llama-era models toward a more complicated world of attention variants, mixture-of-experts routing, multimodal encoders, and multi-GPU inference constraints. The piece compares this trajectory to recommendation systems, where performance optimization became so load-bearing that experimentation itself required optimized baselines. The takeaway is that future ML tooling needs composability built in, with examples like PyTorch FlexAttention offering a way to explore variants without falling off a performance cliff.
Discussion: Mixed — HN readers broadly accepted the premise that modern LLM stacks are becoming more complex, especially around attention variants, MoE, routing, and inference optimization. But several commenters challenged the article’s comparison as cherry-picked or not apples-to-apples, and a large side thread got distracted by accusations and counter-accusations about whether the post was AI-written. (LLM architecture complexity is increasing as scaling gains get harder, Optimization can become a barrier to evaluating new model ideas, Debate over whether the Llama 3 versus Nemotron comparison was fair)
0:00 / 1:03biotech A 15-minute Lyme test for ticks is coming home
A Boston startup plans to sell LymeAlert in August: a $40 at-home test that grinds up removed ticks and uses treated paper to detect Lyme disease bacteria in about 15 minutes. The pitch is faster triage after a bite—helping doctors decide on prompt preventive antibiotics while avoiding unnecessary visits and antibiotic use when a tick is negative. Experts quoted in the article call the idea promising if accurate, but warn that false positives could cause panic and that the test does not detect other tick-borne hazards such as agents linked to Alpha-gal syndrome.
Discussion: Mixed — HN readers were highly engaged, but the discussion quickly moved from the product itself into personal accounts of Lyme, chronic fatigue, long COVID, mold exposure, sleep problems, and other hard-to-diagnose conditions. The mood was cautiously interested in faster tick testing, while also wary of false positives, incomplete pathogen coverage, and internet-driven medical speculation. (Personal stories of suspected or confirmed tick-borne illness, Concern about chronic, ambiguous symptoms and difficult diagnoses, Interest in earlier antibiotics and avoiding unnecessary treatment)
0:00 / 0:39science Slow breathing may change the brain’s appetite for risk
A new Neuron paper reports that slow breathing, especially techniques involving prolonged exhalation, can modulate brain function and shift reward-related behavior, including risk-taking. The work matters because it links a simple bodily intervention to autonomic state and reward processing, with possible implications for anxiety, panic disorder, and depression. The discussion focused less on the mechanics of the study and more on whether making people more reward-seeking is calming, risky, or both.
Discussion: Mixed — HN was broadly intrigued and personally engaged, with many commenters connecting the finding to public speaking, anxiety, yoga, tactical breathing, and beta blockers. The main tension was surprise that parasympathetic activation and prolonged-exhalation breathing appeared to increase reward responsiveness or risk tolerance, which some saw as useful emotional recalibration and others read as counterintuitive or potentially over-framed as beneficial. (Slow breathing as a practical tool for public speaking and acute anxiety, Debate over whether increased risk-taking is good, bad, or context-dependent, Connections to vagal tone, parasympathetic activation, pranayama, and tactical breathing)
0:00 / 0:42science Your Brain Wasn’t Built for the Bad-News Firehose
The article argues that rising news avoidance is not apathy but a predictable response to brains evolved to prioritize local threats being exposed to constant global crises. It cites Reuters Institute survey data on news avoidance, research showing negative words increase news click-through rates, and work on “Problematic News Consumption” linked to distress. The practical advice is not total avoidance, but bounded news windows, fewer higher-quality sources, skepticism toward rage bait, and connecting awareness to concrete action.
Discussion: Mixed — HN broadly accepted the premise that modern feeds exploit threat detection and overwhelm people, but the thread quickly moved into debate over civic responsibility, political agency, and whether unplugging is healthy or escapist. Many commenters emphasized focusing on what you can actually affect, while others pushed back that news still matters for voting, community action, and people directly affected by events. (negativity bias and attention capture, doomscrolling and social media feeds, local agency versus global helplessness)
CSSQuake is a browser-based recreation of Quake where the rendering is done with CSS, with runtime game logic handled in TypeScript. The project hit HN as a novelty programming showcase: less a practical game engine than a demonstration of how far modern CSS and browser rendering can be pushed. Commenters also surfaced the source repository and discussed its relationship to CSS Doom and earlier CSS 3D experiments.
Discussion: Positive — HN is largely delighted by the technical stunt, with many calling it impressive, funny, and smile-inducing. The main skepticism is around performance and the “CSS only” framing, since commenters note the rendering is CSS but the game logic runs in TypeScript/JavaScript. Browser differences, especially Safari/WebKit versus Firefox or Chrome, were a recurring practical theme. (Admiration for the hack and nostalgia for Quake, Performance varies sharply by browser and machine, Debate over whether it is truly CSS-only because JS/TypeScript handles game logic)
0:00 / 0:21software Linux finally removes strncpy after a six-year cleanup
Linux 7.2 has removed the kernel’s strncpy API after about six years and roughly 362 commits spent eliminating all in-kernel users. The function had long been deprecated because its semantics around NUL termination and zero-padding are easy to misuse, and it carried performance costs from redundant zero-filling. 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 exact string or memory-copy intent.
Discussion: Mixed — HN is broadly approving of the removal, with respect for the long, unglamorous engineering work behind it. The mood is also exasperated: commenters use the news as a jumping-off point to criticize C string semantics, NUL-terminated strings, and decades of standard-library inertia. (Praise for slow infrastructure cleanup and bug reduction, Frustration with strncpy’s confusing NUL-termination and padding behavior, Debate over whether C should have adopted fat pointers, ranges, or Pascal-style strings)
0:00 / 0:27software F-15 Strike Eagle II reverse-engineering project calls for DOS test pilots
A hobby project reconstructing the C source code for the 1989 DOS game F-15 Strike Eagle II says it has reached a major milestone: all C code has been reconstructed for all executables, with most assembly-only code replaced by functional C. The team is now asking people with the original 451.03 game plus the Desert Storm expansion to test release v0.9.1 under DOSBox or real DOS and report crashes, graphical glitches, and input bugs. The project is intentionally bug-for-bug for now, so testers are asked to compare against the original before filing issues.
Discussion: Positive — HN is broadly enthusiastic, with many commenters nostalgic for MicroProse-era flight sims and impressed by the reverse-engineering milestone. The main discussion centers on why source reconstruction matters when DOSBox already exists, plus legal questions around reverse engineering and whether AI makes this work easier. (nostalgia for DOS flight simulators, source reconstruction as preservation and modding enabler, legal uncertainty around reverse engineering old games)
0:00 / 0:28software TownSquare makes websites feel inhabited — and immediately hits the moderation wall
TownSquare is a tiny embeddable presence layer for websites: add one snippet and visitors can see each other, move around, interact, and post short messages without accounts. The pitch is a more human, inhabited web, but the HN-driven live demo was quickly flooded with crude spam, turning the launch into an immediate discussion about moderation and abuse resistance. The project’s creator noted that moderation exists but the HN traffic arrived before more keyword rules and safeguards were in place.
Discussion: Mixed — HN liked the playful, lightweight idea of adding live presence to static websites, but the live demo quickly became a case study in anonymous abuse and spam. The discussion was dominated by moderation, rate limiting, filtering, and whether site owners or clients should control what gets shown. (Delight at small, playful web interactions, Concern over anonymous real-time chat abuse, Moderation, spam prevention, and rate limiting)
0:00 / 0:26software Linux Async I/O: Epoll’s Old Guard Meets io_uring
The article compares Linux’s traditional epoll readiness model with io_uring’s newer completion-based model, using the author’s TinyGate reverse proxy project as the motivating example. The core claim is that epoll still requires separate read and write syscalls after readiness notifications, while io_uring can batch submissions and completions through shared ring buffers, reducing syscall overhead. It matters because high-throughput proxies and servers are often limited less by business logic than by kernel crossings, memory movement, and CPU/cache behavior.
Discussion: Mixed — HN liked the practical, educational walkthrough, but the discussion quickly turned from simple epoll-versus-io_uring framing to real-world performance caveats: CPU affinity, allocator choices, NAPI busy polling, zero-copy, framework overhead, and security tradeoffs. Several commenters agreed io_uring can be faster, but warned that integration details, kernel support, and past vulnerabilities make it far from an automatic default. (io_uring can reduce syscall overhead, especially with batching and zero-copy, Performance depends heavily on CPU pinning, cache locality, NIC queues, and avoiding cross-core communication, Existing async frameworks may not expose the parts of io_uring that make it fastest)
0:00 / 0:23software Cloudflare gives AI agents one-hour deploy accounts
Cloudflare announced Temporary Cloudflare Accounts for Agents, letting tools run `wrangler deploy --temporary` to deploy a Worker without first creating or authenticating a normal Cloudflare account. The deployment lives for 60 minutes, can be claimed by a human and made permanent, or expires automatically. The point is to remove browser-based signup, OAuth, MFA, and token-copying steps that break autonomous coding-agent workflows, while still giving agents a deploy-and-verify loop.
Discussion: Mixed — HN liked the low-friction idea, especially as a general-purpose scratch or preview deployment tool, but the discussion quickly shifted to concerns Cloudflare has not solved: hard billing caps, abuse prevention, and what it means for agents to accept terms of service. The overall mood was cautiously interested, with a lot of skepticism about operational and security consequences. (Temporary Workers deployments could be useful beyond AI agents, including PR previews and code review, Strong demand for hard billing caps on Cloudflare Workers, especially if agents can deploy and iterate automatically, Concerns that ephemeral infrastructure could be abused for phishing or malicious hosting)
0:00 / 0:40software X11 windows, floating in Vision Pro space
UHF X11 is a visionOS app that turns Apple Vision Pro into an X11 display server, letting trusted external X clients open as native spatial windows. It supports standard X11 TCP connections, MIT-MAGIC-COOKIE-1 authentication, bitmap font packs, retro CRT-style effects, and experimental indirect GLX. The appeal is both practical and nostalgic: old Unix GUI clients like xterm, xclock, xcalc, and TWM can be placed around a room as Vision Pro windows.
Discussion: Mixed — HN liked the retro-computing hack and the humor of bringing X11, TWM, GLX, and classic visual effects into spatial computing. But much of the thread pivoted to skepticism about Apple Vision Pro itself: price, weight, motion sickness, prescription-lens friction, ecosystem control, and whether more open Linux-friendly headsets will be better developer platforms. (Nostalgia and amusement around X11 in AR, Technical curiosity about GLX, xeyes, Xorg compatibility, and rootless windows, Strong reluctance to buy Vision Pro because of cost and ergonomics)
0:00 / 1:41security Loupe shows what iOS apps can quietly learn about your device
Mysk released Loupe, a free open-source iOS and iPadOS app that demonstrates the device-fingerprinting surface available through public Apple APIs. It shows real raw values on-device, grouped into passive signals, permission-gated data, and more advanced techniques like URL-scheme probing and Keychain persistence across reinstalls. The project’s point is that apps may be able to recognize users across contexts without needing obvious identifiers like a name, email, or location.
Discussion: Negative — The discussion is largely alarmed and critical of Apple’s permission model. Commenters focused less on Loupe as an app and more on the fact that iOS allows many fingerprinting-relevant signals and unrestricted network access by default, with several comparing iOS unfavorably to GrapheneOS, LineageOS, and even China-market iPhones that expose per-app network controls. (Calls for opt-in or toggleable per-app internet access on iOS, Concern that device fingerprinting can work without names, emails, or location, Frustration that users cannot easily verify local-first or offline-first app claims)
0:00 / 0:42hardware MicroVMs come to Proxmox, with container-like boot times
A homelab developer built pve-microvm, a Debian package that adds QEMU’s microvm machine type into Proxmox VE with web UI hooks, a custom Linux kernel, a tiny initrd, OCI image import, and templates for multiple guest OSes. The goal is to get VM-grade KVM isolation with startup behavior closer to containers: stripped-down virtio-only guests that can boot in hundreds of milliseconds, with some examples much faster. It matters because Proxmox users often have to choose between efficient but weaker-isolated LXC containers and heavier full VMs; this project explores a middle ground, though it currently relies on patching Proxmox internals.
Discussion: Mixed — HN is impressed by the Proxmox integration and the promise of fast, isolated guests, but cautious about maintainability and production readiness. The discussion treats microVMs as increasingly relevant, especially for homelabs, coding-agent sandboxes, and lightweight isolated workloads, while noting gaps around GPU/CUDA support, orchestration, and upstream support. (Strong interest in first-class Proxmox microVM support, Tradeoffs versus LXC containers and full VMs, Concern about patching Proxmox internals and upgrade fragility)
0:00 / 0:32startups StartupWiki pitches a free Crunchbase alternative, but HN wants proof
StartupWiki is a new free startup database meant to feel more like Wikipedia than Crunchbase: no subscriptions, no required accounts for browsing, and straightforward company profiles with search, filtering, categorization, and a public API in progress. The HN launch drew interest because startup data is often paywalled or fragmented, especially for early-stage companies. But commenters found missing companies and incorrect or stale data, pushing the project toward the hard parts: provenance, licensing, trust, and repeatable data collection.
Discussion: Mixed — HN liked the premise of a free, uncluttered startup database, but the thread quickly centered on coverage gaps, data accuracy, provenance, licensing, and whether AI-generated profiles can be trusted. The founder was active in replies, fixing bugs and describing plans to add YC companies, source links, an agent ledger, and manual approvals, which helped the tone stay constructive despite skepticism. (Demand for verifiable sources behind “verified” badges, Sparse and inaccurate company data, including outdated funding and runway estimates, Calls for open source, clear content licensing, and regular data dumps)
0:00 / 1:31policy Claude’s ID checks spark privacy backlash
Anthropic’s support page says Claude may prompt users for identity verification for certain capabilities, platform-integrity checks, safety, compliance, or age-related reasons, using Persona as the verification partner. The page says accepted documents are physical government-issued photo IDs from most countries, and that Anthropic uses verification data to confirm identity rather than train models. The HN discussion matters because users are treating AI account KYC as a major trust boundary: not just a support workflow, but a signal about who can access frontier models and who can be tracked or excluded.
Discussion: Negative — HN reaction is strongly hostile to Anthropic’s identity verification, with many commenters saying they would cancel or avoid Claude rather than upload government ID and a selfie to Persona. A recurring corrective note is that the support page and ID process appear to have existed for months, so commenters linking it directly to the latest model-access controversy may be overstating the timing. Even so, the discussion frames KYC for AI as a privacy, surveillance, export-control, and platform-power problem. (Distrust of Persona and third-party ID verification, Fear of government access, surveillance, and legal-process requests, Concern that non-US users may lose access to top US models)
0:00 / 1:46policy A slick AI fan site allegedly copied an entire bestseller
Andy Baio reports that a polished site using a near-identical domain to John Koenig’s The Dictionary of Obscure Sorrows reproduced the book’s foreword and all 311 entries, while replacing the original artwork with DALL-E 2 images and adding a GPT-4-powered “submit a sorrow” feature. Koenig told Baio he had nothing to do with the site; the agency credited in the footer, Qontour, described it in its portfolio as a fan-built Webflow project. The case matters because it blends old-fashioned copyright infringement with AI-era packaging: a copied literary work turned into a showcase for web design and generated content.
Discussion: Negative — HN commenters were broadly angry about the apparent wholesale copying and saw it as a warning sign for creators in an AI-assisted web. The discussion quickly moved from this specific site to frustration with DMCA enforcement, platform incentives, and the sense that small creators face high legal costs while copying gets easier. (Outrage over verbatim copying of creative work, DMCA takedowns seen as ineffective for smaller creators, AI lowers the cost of rebranding or repackaging others’ work)
0:00 / 1:14general Beyond All Reason revives Total Annihilation-style RTS at massive scale
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 HN thread turned into both a celebration of classic RTS design and a cautionary discussion about how hard it can be for casual or new players to enter competitive public lobbies. The story matters because it shows there is still strong appetite for large-scale PC RTS games, but also that community design and onboarding can make or break the experience.
Discussion: Mixed — HN is impressed by the game itself—especially its Total Annihilation lineage, large-scale real-time simulation, performance, and free-to-play/open RTS spirit—but the dominant discussion is a warning about multiplayer culture. Many commenters describe public team lobbies as hostile to newcomers and overly meta-driven, while others say the experience improves in solo, PvE, smaller matches, noob-friendly lobbies, or private groups. (Nostalgia for Total Annihilation and 1990s RTS games, Praise for scale, simulation, controls, and performance, Toxicity and pressure in public 8v8 team lobbies)
0:00 / 0:24general Google says IPv6 has crossed the 50% line
Google’s public measurements show IPv6 reaching 50% of users accessing Google services for the first time. APNIC argues the milestone is significant, but also notes its own weighted global measurement is lower, around 42%, because different methodologies and country-level weighting produce different global views. The bigger takeaway is that IPv6 is now mainstream, but adoption remains highly uneven by country, ISP, access network, and service provider.
Discussion: Mixed — HN treats the 50% mark as a real milestone, but the discussion is dominated by frustration that IPv6 still feels uneven and optional after decades. Commenters repeatedly call out lagging ISPs, IPv4-only services, home-router limitations, and weak incentives, while others argue IPv6 already improves performance where IPv4 is stuck behind CGNAT. (50% is meaningful but measurement-dependent, ISP support remains patchy, with Virgin Media and T-Mobile/Odido singled out, Large regional differences: India, France, Viet Nam, Saudi Arabia, and mobile networks matter)
0:00 / 0:24general Finland’s libraries lend sewing machines—and a lot more
The BBC profiles Finland’s expanding vision of public libraries, where Helsinki’s Oodi and other branches lend or provide access to sewing machines, music studios, 3D printers, sports gear, meeting rooms, and help with digital bureaucracy. The article argues that Finland treats libraries less as book warehouses and more as democratic infrastructure: open, tax-funded public spaces that support inclusion, skills, debate, and everyday problem-solving. Finland’s spending is cited at nearly €371 million in 2025, or €65.78 per person, compared with lower per-person figures cited for the UK and US.
Discussion: Positive — HN was broadly enthusiastic, with many commenters sharing examples of their own local “libraries of things” lending tools, instruments, makerspace gear, museum passes, and other non-book resources. The main reservations were practical: maintenance burden for complex equipment, long waitlists when inventory is thin, and concerns that some libraries are underfunded or strained by broader social problems. (Libraries as third places and civic infrastructure, Tool and equipment lending as practical public value, Makerspaces, 3D printers, sewing machines, and creative access)
0:00 / 0:23general SMPTE opens its media standards library
SMPTE says its full catalog of published standards, recommended practices, engineering guidelines, registered disclosure documents, and future releases is now freely accessible. The group frames the move as a way to improve media-industry interoperability, especially as workflows shift toward IP, AI authenticity, and content provenance. SMPTE also says it is modernizing standards work with GitHub-based workflows, structured HTML authoring, and an integrated publishing pipeline.
Discussion: Positive — HN is strongly supportive of SMPTE making its standards free, seeing it as overdue and likely to improve interoperability and implementation. The thread quickly broadens into frustration with paywalled standards from ISO, ANSI, IEEE, construction-code bodies, and others, with some commenters noting that standards development still has real overhead costs that need funding. (Open standards should be freely accessible, Paywalled specifications create barriers for hobbyists, developers, and implementers, Standards bodies need sustainable funding despite volunteer technical work)
0:00 / 0:43general HN Marks Father’s Day With Hacker Dad Stories
A short Tell HN post wished the community a happy Father’s Day through a personal memory: in Soviet-controlled Poland, the poster’s uncle filled a father role, bringing them into hands-on hacking with rocket cars and even wiring a stolen Milicja siren onto a banana bike. The post resonated because it framed technical curiosity as something often inherited through family, mentors, and improvised childhood adventures. The comments turned into a low-key tribute thread for fathers, father figures, and the messy, joyful work of raising kids.
Discussion: Positive — The thread is overwhelmingly warm and reflective, with commenters sharing memories of fathers, uncle-as-father figures, first Father’s Day moments, and the way parenting reshaped their priorities. A few side discussions add practical dad-life advice and note that Father’s Day is not celebrated on the same date everywhere, but the overall mood stays affectionate and celebratory. (gratitude for fathers and father figures, tinkering and technical curiosity passed across generations, new parent joy and exhaustion)