0:00 / 1:17ai A 744-billion-parameter open model squeezes onto local hardware
Unsloth published instructions and GGUF quantizations for running Z.ai’s GLM-5.2 locally, a 744B-parameter mixture-of-experts model with 40B active parameters and a claimed 1M-token context window. The guide says a 2-bit dynamic quantization takes 239GB of disk space and can fit on high-memory local systems, while smaller quants make the model dramatically more accessible than the full 1.5TB version. The significance is less that this is easy for normal laptops, and more that frontier-adjacent open models are moving into llama.cpp-style local workflows for hobbyists, privacy-sensitive users, and small teams.
Discussion: Mixed — HN is impressed that GLM-5.2 can be packaged as GGUF and run through local tooling, but the excitement is heavily qualified by hardware cost, speed, quantization quality, and power economics. Commenters see real value for privacy, control, experimentation, and some company workloads, while many argue cloud APIs remain faster and cheaper for most users today. (Local inference is becoming more practical, but still not mainstream, Hardware requirements remain steep: hundreds of GB of RAM and/or expensive GPU setups, Privacy, sovereignty, and avoiding cloud model downgrades are major motivations)
0:00 / 1:20ai FUTO ships a new swipe-typing model trained on a million user swipes
FUTO says it collected more than 1 million voluntary English QWERTY swipe gestures starting in August 2024, filtered the data, released the dataset under MIT on Hugging Face in March 2025, and used it to train and evaluate new swipe-typing models. The launch matters because high-quality swipe input has largely been dominated by incumbent mobile keyboards, and FUTO is positioning this as a more private alternative with an openly released dataset.
Discussion: Mixed — HN is broadly excited: many commenters say the new swipe model makes FUTO Keyboard feel close to Gboard, and they like the privacy-oriented alternative to Google and Microsoft keyboards. The mood is tempered by complaints about licensing complexity, remaining accuracy issues, capitalization and context prediction bugs, and gaps around iOS and multilingual support. (Strong demand for a private, non-Google swipe keyboard, Accuracy now feels close to Gboard for some users, Interest in layouts optimized specifically for swiping, including ClearFlow)
0:00 / 0:26ai Baidu’s Unlimited-OCR Tackles Long PDFs Without Blowing Up VRAM
Baidu has released Unlimited-OCR, a GitHub project and newly posted arXiv paper aimed at one-shot, long-horizon OCR parsing, with the model also available on ModelScope. The repo describes inference via Hugging Face transformers on NVIDIA GPUs, SGLang serving, an OpenAI-compatible streaming API, and batch processing for image directories or PDFs. The work positions itself as an extension of ideas from Deepseek-OCR, Deepseek-OCR-2, and PaddleOCR, and matters because long-document OCR often runs into context and memory limits that force brittle page-splitting pipelines.
Discussion: Positive — HN was broadly interested and technically engaged, with commenters seeing the project as a practical answer to long-document OCR memory limits. The main discussion centered on whether sliding-window-style approaches are novel here, when page slicing is still sufficient, and where extra context helps with messy scans, tables, and structured documents. A large side thread used the OCR launch as a springboard into optical music recognition, where commenters saw major unmet potential for AI. (long-context memory and KV-cache limits, sliding-window attention versus page-by-page OCR, practical document slicing and overlap strategies)
0:00 / 0:23ai Mistral ships OCR 4 with boxes, confidence scores, and self-hosting
Mistral released OCR 4, a document-understanding model that extracts text along with bounding boxes, block types, and word-level confidence scores. It supports 170 languages, can run in a single self-hosted container for enterprise customers, and is positioned as an ingestion layer for search, RAG, compliance, and document workflows. Mistral says it leads tested systems in human preference evaluations and public benchmarks, while also acknowledging benchmark limitations; API pricing is $4 per 1,000 pages, or $2 with batch discounts.
Discussion: Mixed — HN was interested in the practical value and low price, but skeptical of Mistral’s benchmark framing and past OCR claims. The discussion drifted heavily into postal-service OCR and address-routing history, with side debates about Mistral’s U.S. presence and European tech commercialization. (Skepticism about internal benchmarks and chart presentation, Interest in pricing and self-hosted deployment, Comparisons with USPS, census, and older OCR systems)
0:00 / 0:26ai Coding agents are entering the outer-loop era
Armin Ronacher argues that coding-agent workflows are shifting from direct prompting to “harness-level loops,” where tasks are queued, retried, recontextualized, or handed between machines after the model would normally stop. He says these loops already work well for porting, benchmarking, security scanning, research, and other mechanically checkable or short-lived artifacts, but he has not had much success using them on code he deeply cares about. The concern is that today’s models tend to amplify local, defensive fixes—fallbacks, null checks, duplicated abstractions—rather than producing simple systems with strong invariants and code humans can still explain.
Discussion: Mixed — HN largely agreed with Ronacher’s caution: agent loops can be powerful, but only where goals are clear, outputs are disposable, or correctness is mechanically verifiable. Commenters were skeptical of handing long-lived, judgment-heavy codebases to autonomous loops, citing loss of understanding, bloated defensive code, weak reviews, and security risk, while several reported real productivity gains when they supplied strong specs and stayed in the loop. (Agents work best with clear specifications and verifiable outcomes, Concern about maintainability, comprehension, and code review becoming rubber stamps, LLMs overproduce defensive fallbacks instead of enforcing invariants)
David Rosenthal’s post argues that major AI platforms have built demand by heavily subsidizing usage, and that the shift from flat subscriptions to token-based pricing is exposing how expensive serious usage can be. Citing Ed Zitron, SemiAnalysis, and business-press reports, it highlights claims that $200-per-month individual plans can consume thousands of dollars of tokens, that OpenAI’s 2025 costs and losses were enormous, and that companies are starting to rein in employee AI use. The broader implication is that enterprise AI adoption may be more price-sensitive than the hype suggests, especially if providers need to raise prices to cover compute, training, and capital costs.
Discussion: Mixed — HN is broadly skeptical that today’s AI pricing and spending levels are sustainable, but commenters split on why: some see a subsidy-fueled bubble and looming crash, while others argue costs will fall quickly or that ROI, not raw token price, is the real issue. The strongest real-world signal in the thread is companies suddenly adding controls, budgets, and escalation around model usage after moving to token-based pricing. (enterprise AI budgets tightening, token-based pricing shock, uncertain inference margins)
0:00 / 0:44ai Anthropic brings Claude into Slack as a team agent
Anthropic launched Claude Tag, a beta Slack integration for Claude Enterprise and Team customers that lets users mention @Claude in channels, connect it to approved tools and data, and delegate asynchronous work. The product is positioned as a multiplayer evolution of Claude Code: one channel-scoped Claude can build context over time, optionally act proactively, and keep its memories and permissions scoped by administrators. Anthropic says its internal version now creates 65% of its product team’s code, a claim that drew both attention and skepticism.
Discussion: Mixed — HN was interested in the multiplayer Slack-agent model, but the discussion leaned skeptical. Commenters worried about token costs, lock-in, permissions, auditability, memory quality, and Anthropic’s claim that 65% of its product team code is created by an internal version of Claude Tag. (Token spend and API billing concerns, Enterprise permissions, compliance, and audit logs, Shared agent context versus privacy boundaries)
0:00 / 1:04biotech Vitamin D may not be magic, but it may not be worthless either
The article argues that the backlash against vitamin D supplements has gone too far: randomized trials undermine the old panacea claims, but may still provide weak positive evidence for people with low-ish vitamin D levels. It walks through the biology of vitamin D, the history of correlations between sunlight and health outcomes, and the limits of using broad trial results to decide what low-level or seasonally deficient people should do. The practical takeaway is cautious: vitamin D is not a miracle supplement, but for people with low levels, supplementation may be a reasonable bet under uncertainty.
Discussion: Mixed — HN readers mostly appreciated the article’s balanced stance, but the discussion was split between skepticism of supplement hype and personal or regional experiences suggesting supplementation helps, especially in winter or deficiency. A recurring theme was that vitamin D blood levels may be a proxy for sunlight, outdoor activity, latitude, diet, skin pigmentation, or broader lifestyle rather than a simple causal lever. (skepticism of supplement hype, sunlight versus supplementation, deficiency thresholds and blood testing)
0:00 / 1:13software F3 proposes a future-proof columnar file format with embedded Wasm decoders
F3 is a research-prototype data file format aimed at improving on columnar formats like Parquet, with a layout designed for efficiency, interoperability, and extensibility. Its most distinctive feature is embedding WebAssembly decoders inside self-describing files, so new encodings can still be read when native decoders are unavailable. The project explicitly warns it is not production-ready and has only been tested on an Intel machine running Debian 12.
Discussion: Mixed — HN was intrigued by the core idea, especially using embedded Wasm decoders to make files forward-compatible across platforms, but the discussion was heavily skeptical about adoption and security. Many commenters argued that Parquet’s biggest advantage is not technical perfection but ecosystem compatibility, and several felt the project’s README did not clearly explain the “why.” (Parquet compatibility and ecosystem lock-in, Embedded Wasm decoders as forward-compatibility mechanism, Security risks of executing code from data files)
0:00 / 1:17software Google fires engineer after viral Workspace CLI
Former Google employee Justin Poehnelt says he was fired two months ago after creating the Google Workspace CLI, which went viral, topped Hacker News, gained thousands of GitHub stars, and attracted many users within days. He says the project drew both leadership interest and legal scrutiny over Google branding, and he believes Workspace leaders were worried about disruption from agents; he also notes Google announced an official Workspace CLI two days before his termination. The story matters because it sits at the fault line between employee-led open-source initiative and the brand, legal, and product-control machinery of a trillion-dollar company.
Discussion: Mixed — HN is split between sympathy for a builder who made a useful, popular tool and skepticism that an employee could release an official-looking Google Workspace project without serious consequences. Many commenters see firing as an extreme bureaucratic response, while others argue Google’s branding, legal, launch, and open-source approval processes make this a predictable outcome. (Tension between hacker culture and big-company process, Use of Google branding and official-looking repositories, Open-source approval and legal review at large companies)
0:00 / 1:26software Accidental wigglegrams hiding in an iCloud photo library
The author discovered that years of taking multiple near-identical photos from slightly different angles had accidentally created the raw material for wigglegrams: looping stereo-like animations. They wrote a script using perceptual hashing to scan an iCloud photo library, find visually similar image runs via Hamming distance, and stitch them into GIF-like wiggles. It matters as a charming example of computational photography applied to personal archives, and as a reminder that “duplicates” can contain spatial and temporal information worth preserving.
Discussion: Positive — HN largely enjoyed the post as a playful, clever hack: using perceptual hashing to mine years of near-duplicate photos for accidental stereo animations. The technical discussion focused on why some wigglegrams work better than others—alignment, horizontal parallax, consistent step size, and subject stability—while a noticeable minority found the constant motion dizzying or migraine-inducing. (Perceptual hashing for finding similar photo runs, Stereo photography history and multi-lens cameras, Image alignment, parallax, and wigglegram quality)
0:00 / 0:22software Remembering Tony Krueger, the engineer behind Word’s red squiggles
Raymond Chen memorialized Tony Krueger, a longtime Microsoft Word engineer whose work made spell checking non-blocking and introduced the now-familiar red squiggles under suspected misspellings, later joined by green grammar markers. The post also notes Krueger’s Windows port of Chip’s Challenge, reportedly done by reverse-engineering the MS-DOS version without source code. The story matters because it highlights a nearly invisible piece of user-interface history that became standard across word processors, editors, and many other apps.
Discussion: Mixed — HN was broadly appreciative of Raymond Chen’s remembrance and of the red-squiggle UI pattern, but the thread quickly mixed tribute with practical complaints about spellcheck in multilingual writing, distraction, autocorrect, and attribution. Several commenters also dug into prior art and documentation questions around real-time spell checking and Krueger’s Chip’s Challenge port. (Respect for overlooked software contributors, Red squiggles as a durable and intuitive UI convention, Annoyance with spellcheck in multilingual or technical writing)
0:00 / 0:22software A visual TikZ editor for LaTeX figures gets Hacker News excited
A developer released TikZ Editor, a free and open-source WYSIWYG editor for TikZ diagrams in LaTeX, available on the web and as a desktop app. It keeps source code and the rendered figure in sync, lets users drag, resize, align, and edit objects visually, and aims to preserve existing formatting rather than rewriting files wholesale. The project matters because TikZ is powerful but notoriously tedious to edit by hand, especially for academic figures where small layout changes often mean repeated coordinate tweaking and recompilation.
Discussion: Positive — The discussion is strongly favorable, with many academics and LaTeX users saying this solves a long-standing pain point: hand-tuning TikZ coordinates. The main reservations are about whether the generated or edited TikZ remains idiomatic, especially around absolute coordinates versus relative structure, plus feature requests for Typst/Cetz, pgfplots, CircuitiTikZ, and presets. (Relief at avoiding manual TikZ coordinate tweaking, Praise for preserving existing source formatting while editing visually, Concern about absolute-coordinate-heavy output and preserving idiomatic TikZ)
0:00 / 0:36software Apple brings Swift Package Index in-house
Swift Package Index, the community-run search and metadata index for Swift packages, has joined Apple. The announcement says SPI will keep operating, remain open source, continue multi-platform package testing, and evolve toward a comprehensive Swift package registry with future work around package signing and identity. This matters because SPI has become core infrastructure for Swift developers, with more than 10,000 packages indexed and millions of compatibility builds processed across Apple platforms, Linux, WebAssembly, Android, and more.
Discussion: Mixed — HN is broadly happy for the Swift Package Index team and sees value in Apple funding a central Swift package registry, but the thread is cautious-to-skeptical about Apple stewardship. The biggest worries are Apple’s open-source track record, developer identity requirements, possible package curation or lock-in, and whether non-GitHub or non-Apple-platform Swift users will be well served. (Congratulations to Dave Verwer and the SPI team, Hope for a more robust Swift package registry, Concern about Apple’s open-source and developer-services history)
0:00 / 0:45security Vulnerability reports hit the LLM spam era
Filippo Valsorda argues that vulnerability reports are no longer automatically “special” because LLMs have made finding plausible bugs cheap and widespread. The scarce work has shifted from discovery and confidential coordination to triage, impact assessment, rapid remediation, and prevention. The piece does not say all reports are worthless; it suggests teams now need to distinguish exceptional, trusted, high-severity reports from the growing mass of noisy ones.
Discussion: Mixed — Commenters largely agreed that vulnerability disclosure inboxes are being flooded with low-quality or AI-generated reports, creating real triage fatigue and making serious issues harder to identify. The mood was frustrated but not purely pessimistic: some saw this as a temporary shakeout that could push teams toward better CI scanning, safer languages, fuzzing, and stronger engineering practices. (AI-generated vulnerability report spam, CVE and dependency-alert fatigue, triage burden for maintainers and security teams)
0:00 / 0:23security Meta pauses employee-tracking tool after internal data exposure
Meta has paused its Model Compatibility Initiative, an internal tool rolled out to US employees in April that collected mouse movements, click locations, keystrokes, and screen content to help train AI systems to use software like humans. WIRED reports an internal security notice said databases containing MCI-derived data were exposed to more Meta employees than intended; Meta says it has no indication the data was improperly accessed and will not re-enable the tool until it is confident in its controls. The story matters because it combines two highly sensitive issues: workplace monitoring and the use of employee activity as AI training data.
Discussion: Negative — HN reaction is strongly negative, with many commenters treating the incident as confirmation of long-running distrust toward Meta’s privacy culture and employee surveillance practices. The thread spends less time on the specific security mechanics and more on Meta’s broader social impact, ethics of working there, and whether its open-source work offsets criticism. (employee surveillance and workplace privacy, Meta’s reputation and ethics, AI training data collection from workers)
0:00 / 0:36security OpenAI launches GPT-5.5-Cyber, but HN wants to know who gets access
OpenAI announced an expansion of its Daybreak cybersecurity effort, including an updated Codex Security plugin, a limited-release full version of GPT-5.5-Cyber for trusted defenders, a partner program, and a Patch the Planet initiative with Trail of Bits and others. The company says GPT-5.5-Cyber reaches 85.6% on CyberGym versus 81.8% for GPT-5.5, and frames the shift as moving from vulnerability discovery to end-to-end patch automation. The big issue for the HN crowd is not the benchmark, but access: many commenters argue the most capable defensive tools are being reserved for selected organizations while ordinary paying users remain locked out.
Discussion: Negative — HN reaction is dominated by frustration over restricted access to frontier cyber models. Commenters see OpenAI’s “trusted defenders” approach as gatekeeping, potentially creating a two-tier security world where large or approved organizations get better protection while smaller developers and non-U.S. users are left with weaker tools. A minority defend limited release on dual-use safety grounds, but the thread is largely skeptical of private companies deciding who counts as a legitimate defender. (restricted access and KYC, dual-use cyber safety, two-tier AI security)
0:00 / 1:18hardware Valve’s new Steam Machine is here, and Hacker News is arguing about the price
Valve has opened reservations for the new Steam Machine, positioning it as a living-room SteamOS PC rather than a traditional console. The launch uses a randomized reservation and waitlist system intended to reduce botting and reseller pressure, with signups open for several days before randomization. The big issue for HN is pricing: commenters repeatedly cite a base price around $1,049 without a controller, and debate whether openness and Steam-library access justify the premium over consoles.
Discussion: Mixed — HN is broadly supportive of Valve’s philosophy—open hardware, SteamOS, Linux gaming, and an anti-scalper reservation lottery—but sharply divided on whether the Steam Machine makes sense at its launch price. Many commenters think the hardware is poor value against consoles like the PS5 Pro, while others argue small-form-factor PC pricing, Steam library reuse, and a polished couch-gaming experience make it defensible. (Price-to-performance concerns versus consoles, Praise for an open PC that can run other apps or operating systems, Reservation lottery seen as a fairer anti-scalping approach)
0:00 / 0:33hardware Gaussian splats are becoming physical objects
Artist Dany Bittel showed a physical print of a Gaussian-splat insect, produced by Crysta AI by voxelizing the splat into a transparent, ink-mixed volume and printing it layer by layer. The result is a glassy, amber-like bee with visible depth and translucency, though Bittel notes some darkness, clumpiness in fine fur, and remaining splat artifacts. It matters because it points to a new workflow for turning photogrammetry-style neural/3D captures into tangible full-color objects, not just screen-rendered scenes.
Discussion: Positive — HN reaction was strongly enthusiastic, with many commenters surprised that full-color, transparent 3D printing could reach this fidelity. The discussion quickly turned practical: what printer/process is involved, how much it costs, whether this becomes a gift-shop product, and why voxelizing splats is the natural bridge to resin-style printing. (Amazement at print fidelity and realism, Curiosity about full-color resin or UV inkjet printing, Speculation on cost, availability, and consumer adoption)
0:00 / 0:37policy America’s giant trucks are making pedestrian crashes deadlier
A New York Times investigation argues that the U.S. shift from sedans to taller pickups and SUVs has made pedestrian crashes more deadly, contributing to a reversal in decades of safety gains. The Times estimates that if vehicles had stayed roughly the same size, about 200 to 400 pedestrians per year would have survived — around 10 percent of the recent increase in pedestrian deaths. The piece focuses on taller hoods, larger blind zones, and the limits of automated pedestrian-detection systems, while automakers and regulators emphasize road design and collision-avoidance technology.
Discussion: Mixed — HN broadly agrees that oversized pickups and SUVs are a serious pedestrian-safety problem, but many commenters push back on treating vehicle size as the main explanation. The discussion is critical of automaker incentives, weak regulation, poor visibility, and U.S. road design, while repeatedly pointing to smartphones, distracted driving, CAFE loopholes, and enforcement as missing or underweighted factors. (Vehicle size and hood height increase pedestrian lethality, Blind zones and poor outward visibility in modern trucks and SUVs, Regulatory incentives, CAFE standards, and automaker profit motives)
0:00 / 0:29policy MSG tracked facial-recognition critics in an internal dossier
404 Media reports that Madison Square Garden compiled an internal document listing activists who publicly criticized its use of facial recognition, including their tweets and comments. The document surfaced in a 45GB cache of MSG data stolen by hackers and posted online. The story matters because MSG has already used facial recognition to block people from venues, raising questions about biometric surveillance, retaliation, and how private businesses handle sensitive watchlists.
Discussion: Negative — HN commenters were broadly alarmed by Madison Square Garden tracking critics and using facial recognition to exclude people, with many calling it dystopian or an abuse of power. The debate split on whether private venues should be free to ban whomever they want, but even some defenders of venue discretion saw blacklisting vocal critics as hard to justify. (biometric surveillance and privacy, private venue power versus public-accommodation norms, MSG tax exemptions and public subsidies)
0:00 / 0:33policy Digital euro advances as Europe targets U.S. card dependence
The digital euro has reportedly cleared an important European Parliament hurdle, moving the European Central Bank’s proposed central-bank digital payment system closer to reality. The political driver is not just modernization: Europe wants a payments infrastructure less dependent on U.S.-based card networks such as Visa and Mastercard. For consumers, the key unresolved question is whether it will feel like a better everyday payment product—or simply a more tightly regulated version of bank-to-bank payments.
Discussion: Mixed — HN commenters broadly understand the strategic sovereignty argument, but are skeptical about whether a digital euro solves everyday payment problems better than existing bank transfers, debit systems, Wero, or national instant-payment schemes. The discussion is split between people who see Visa and Mastercard as expensive foreign-controlled rails worth replacing, and people worried about KYC, surveillance, loss of credit-card-style fraud protection, and yet another complex EU payment layer. (European payment sovereignty versus dependence on Visa and Mastercard, Confusion over credit cards, debit cards, charge cards, and country-by-country European differences, Consumer protections such as chargebacks, fraud handling, and purchase insurance)
0:00 / 0:50general Jerry’s Map: a lifelong imaginary city built one card at a time
Jerry began drawing an imaginary city in 1963, set it aside in 1983, then resumed after his son rediscovered it in the attic. The project has grown into more than 4,000 eight-by-ten-inch panels arranged in a coordinate grid, revised through a custom deck of roughly 100 instruction cards. Its appeal is less about a finished map than a living system: chance, rules, and decades of manual work producing a sprawling two-dimensional world.
Discussion: Positive — HN reacted with warm fascination and nostalgia. Commenters admired the decades-long commitment, the rule-driven creative process, and the handmade, meditative quality of the work; many shared childhood memories of drawing imaginary maps or compared it to outsider art, tabletop RPGs, Dwarf Fortress, Nomic, and procedural worldbuilding. A small thread pushed back against turning the idea into an AI-generation prompt, arguing that the human slowness and process are the point. (nostalgia for childhood mapmaking, admiration for long-term outsider art, interest in rule-based creativity and randomness)
0:00 / 0:26general Bunny makes its DNS free, with one small catch
Bunny.net says Bunny DNS will no longer charge per DNS query and now includes DNS hosting for up to 500 domains per account, with no query limits and features like smart records and health monitoring included. The company frames DNS as the entry point to the rest of its platform, tying it to CDN acceleration, edge security, DNSSEC using NSEC Black Lies, IPv6 support, and newer record types like HTTPS, SVCB, TLSA, CDS, and CDNSKEY. The catch is that Bunny accounts are still subject to the platform’s standard $1/month minimum spend, so the HN debate quickly focused on whether “free” is the right word.
Discussion: Mixed — HN was broadly favorable toward Bunny and several users praised its CDN, pricing, speed, and appeal as a European alternative to Cloudflare. The main skepticism centered on the $1/month account minimum, with some commenters arguing that calling DNS “free” is misleading, plus a few operational cautions about zone import/export reliability. A large side thread drifted into Hetzner price increases and the competitiveness of European infrastructure providers. (Praise for Bunny as a fast, low-cost Cloudflare alternative, Interest in EU-based infrastructure providers, Debate over whether the $1/month minimum undermines the “free DNS” claim)
0:00 / 0:37general The 23-Foot Banana Car Keeps Getting Pulled Over
Steve Braithwaite, owner and builder of the 23-foot Big Banana Car, says he has been pulled over hundreds of times in 15 years and more than 250,000 miles of driving. The latest stop in Billings, Montana, reportedly involved a license-plate issue and ended without a ticket, but Braithwaite says many past stops were really about officers wanting photos, jokes, or a closer look. The car began as a 2008 joke inspired by oddball custom vehicles on Top Gear and has become Braithwaite’s daily driver and the centerpiece of his planned “World Needs More Whimsy Grand Tour.”
Discussion: Mixed — HN enjoyed the absurdity of a street-legal giant banana and piled on puns, sightings, and art-car trivia, but the thread quickly split over whether police stops for photos and jokes are harmless fun or an abuse of authority. A minority pushed back that the driver seems to take it in stride and that the story is meant as whimsy, while others argued any non-consensual traffic stop is coercive even when the target is a novelty vehicle. (banana puns and pop-culture references, appreciation for art cars and roadside whimsy, questions about custom-vehicle registration and chassis details)