0:00 / 1:08ai Study Finds GPT-5.5 Hallucinates Far More Than Open-Source GLM-5.2
A new analysis claims GPT-5.5 produces hallucinated answers nearly three times more often than the open-source, MIT-licensed GLM-5.2 model, despite GPT-5.5's much larger parameter count. The report argues that scaling up model size and training data is yielding diminishing returns, with accuracy plateauing and hallucination rates sometimes worsening. The findings challenge the assumption that bigger models are always better and highlight the need to focus on truthfulness and uncertainty calibration, not just benchmark scores. This could impact how labs and users evaluate and select AI models going forward.
Discussion: Mixed — HN commenters are divided: some question the methodology and generalizations about model size and hallucination, while others agree that bigger models aren't always better and that training data quality and evaluation methods are crucial. There's broad agreement that hallucination is a nuanced, multifactorial problem not solved by scaling alone. (Skepticism about claims that bigger models hallucinate more, Importance of training data quality and methods, Benchmarks vs. real-world performance)
0:00 / 0:22ai Cloudflare Rolls Out 60-Minute Temporary Accounts for AI Agent Deployments
Cloudflare has introduced Temporary Accounts for AI agents, allowing code to be deployed instantly via the Wrangler CLI without needing to sign up or authenticate first. These ephemeral accounts last 60 minutes and can be claimed for permanent use, streamlining agent-driven workflows and rapid iteration. The move aims to reduce friction for automated deployments, but raises questions about abuse prevention and billing safeguards. The feature is positioned as part of Cloudflare's broader push for frictionless, agent-ready developer experiences.
Discussion: Mixed — Commenters are excited about the frictionless deployment and preview capabilities, but express significant concerns about security, abuse potential, and the absence of hard billing caps. There's a blend of enthusiasm for the developer experience and skepticism about Cloudflare's business motives and risk mitigation. (Frictionless deployment, Potential for abuse/malicious use, Lack of hard billing caps)
0:00 / 0:23ai Claude Rolls Out ID Verification, Sparking Privacy Backlash
Anthropic is introducing identity verification for some Claude AI users, requiring government-issued photo IDs through its partner Persona Identities. The move aims to prevent abuse, enforce policies, and meet legal obligations, but has triggered strong concerns over privacy, surveillance, and access—especially for non-US users. Many see this as a response to government pressure and recent export controls, with parallels drawn to similar measures by OpenAI.
Discussion: Negative — The HN discussion is overwhelmingly negative, with users expressing deep concerns about privacy, surveillance, government overreach, and the exclusion of non-US users. Many see the move as ineffective against bad actors but harmful to regular users, and some threaten to cancel subscriptions or switch providers. (privacy and surveillance fears, government and corporate overreach, exclusion of non-US users)
0:00 / 0:37ai LLMs Get More Complex: From Clean Stacks to Complicated Architectures
A recent blog post highlights how large language models (LLMs) have evolved from relatively simple, repeated Transformer stacks (like early Llama) to far more intricate architectures. Modern LLMs now incorporate advanced attention variants, mixture-of-experts routing, integrated vision/audio, and multi-GPU inference, mirroring the complexity seen in recommendation systems. This increased complexity is driven by the need for incremental performance gains and efficient hardware utilization, making research and optimization much harder.
Discussion: Mixed — HN commenters are split: some agree that LLMs are getting more complex and that this mirrors the history of other ML systems, while others argue the changes are overstated or poorly illustrated by the blog's comparison. The discussion also veers into debates about AI-generated writing and the challenges of understanding modern ML research. (LLM architectural complexity, Optimization vs. innovation, Comparisons between model generations)
0:00 / 0:40biotech Slow Breathing Shown to Influence Brain Function and Risk-Taking
A new study published in Neuron finds that slow, controlled breathing can modulate brain function and increase risk-taking behavior by enhancing cardiac parasympathetic activity. This research challenges the common assumption that slow breathing only promotes calmness and caution, suggesting it may also shift reward processing and risk tolerance. The findings have potential implications for anxiety, depression, and performance under stress.
Discussion: Mixed — Commenters are intrigued by the counterintuitive finding that slow breathing may increase risk-taking, with some expressing skepticism or concern about the framing of the results. Others share personal experiences and practical uses of breathing techniques for anxiety and performance, while a few debate the broader psychological and cultural implications. (unexpected effects of slow breathing, risk-taking vs. calmness, applications in public speaking and stress)
0:00 / 0:42biotech Brains Overwhelmed by Constant Bad News, Study Finds
A new article highlights how the human brain, wired for local threats and negativity bias, is struggling to process today's relentless global news cycle. Studies show that exposure to nonstop negative news can cause psychological distress, leading to avoidance behaviors and even clinical symptoms. The piece argues that news fatigue is a natural response to an environment our brains weren't designed for, and suggests strategies like limiting news intake and focusing on actionable information to protect mental health. This matters as news avoidance rises worldwide, potentially impacting civic engagement and well-being.
Discussion: Mixed — HN commenters are divided: some advocate for deliberate news avoidance to preserve mental health, while others stress the importance of staying informed for civic duty. Many recognize the psychological toll of constant bad news, but debate how much agency individuals have to effect change and whether disengagement is responsible. There's also skepticism about performative activism and the actual impact of personal actions on global issues. (news fatigue and mental health, negativity bias in media, civic responsibility vs. self-care)
0:00 / 0:55software Quake Recreated in CSS: A Playable Browser Demo
A developer has recreated the classic Quake game using CSS for rendering, with game logic handled by TypeScript, resulting in a playable demo directly in modern browsers. The project, CSSQuake, showcases the surprising versatility of CSS and web technologies, though performance varies by browser and hardware. The source code is open on GitHub, and the project has sparked comparisons to similar CSS-based game engines. This experiment highlights both the power and the limitations of pushing web standards beyond their intended use.
Discussion: Positive — The HN discussion is overwhelmingly positive, with users expressing amazement at the technical feat and nostalgia for the original Quake. Some note performance issues and browser differences, but these are discussed in a constructive, often humorous tone. The project is praised as both a technical showcase and a playful demonstration of web technology limits. (Technical achievement, Web technology creativity, Browser performance)
0:00 / 0:18software Classic DOS Flight Sim Reborn: F-15 Strike Eagle II Needs Test Pilots
A community-driven reverse engineering project has successfully reconstructed the C source code for the 1989 DOS game F-15 Strike Eagle II, moving all data and most routines from assembly to C. The project is now seeking volunteers to test the reconstructed game for bugs and glitches, with the goal of achieving a bug-for-bug faithful recreation before future ports to modern platforms. This milestone not only preserves gaming history but also opens the door to modding and further enhancements.
Discussion: Positive — The HN discussion is enthusiastic and supportive, with many expressing nostalgia, technical curiosity, and admiration for the project's progress. Commenters highlight the value of source-level access for modding and preservation, and share insights on reverse engineering, legalities, and the role of AI. Some questions about the necessity versus emulation are addressed constructively, and the overall tone is encouraging. (nostalgia and preservation, technical challenge and learning, open source and modding potential)
0:00 / 0:18software Linux Kernel Finally Drops Buggy strncpy After Six-Year Effort
The Linux kernel has removed the strncpy() API in version 7.2, ending a six-year campaign to eliminate this problematic function. Known for causing subtle bugs and performance issues due to confusing semantics and inefficient zero-padding, strncpy has been replaced by safer, more explicit alternatives. This change required over 360 patches and significant coordination across the kernel's vast codebase. The move is seen as a major step toward improving kernel reliability and code clarity.
Discussion: Mixed — Comments reflect both praise for the long-overdue removal of a problematic API and frustration with C's historical design decisions. Some applaud the grind of maintaining and improving core infrastructure, while others lament the convoluted state of C string handling and slow pace of progress. (legacy C API problems, systems engineering challenges, string handling in C)
0:00 / 0:18software Epoll vs. io_uring: The Future of Async I/O on Linux
A developer recounts the evolution of an educational reverse proxy project, moving from synchronous I/O to epoll, and ultimately to io_uring for better performance. The article explains the architectural differences: epoll uses a readiness model with more syscalls, while io_uring offers a completion model with fewer syscalls and better batching, leading to significant performance gains on modern Linux kernels. The HN discussion dives deep into practical tuning, security concerns, and real-world benchmarks, highlighting both the promise and current limitations of io_uring.
Discussion: Mixed — Discussion is enthusiastic about io_uring's technical potential and performance, but tempered by concerns over security, real-world integration, and the complexity of tuning for optimal results. Many share practical experiences and nuanced takes on when each approach is best. (Performance tuning and benchmarking, Security and kernel support, Architectural trade-offs)
0:00 / 0:36software X11 Comes to Apple Vision Pro with UHF X11
A new project, UHF X11, brings the classic X11 windowing system to visionOS and the Apple Vision Pro headset. This allows users to run X11 applications in spatial windows, complete with retro effects and support for external X11 clients. The project highlights both the enduring relevance of X11 and the creative possibilities of blending vintage computing with cutting-edge AR hardware.
Discussion: Mixed — Comments are enthusiastic about the technical achievement and X11's resilience, but skeptical about the Vision Pro's price, comfort, and closed ecosystem. There's also discussion of alternatives like the upcoming Steam Frame and broader issues with AR headsets. (X11's longevity, Vision Pro's limitations (price, comfort)
0:00 / 1:42security Loupe App Reveals How iOS Apps Fingerprint Your Device
Loupe is a free, open-source iOS and iPadOS app that exposes what device data native apps can access, demonstrating how seemingly innocuous readings can be combined for device fingerprinting. By showing raw values from public APIs, Loupe raises user awareness about privacy risks and the extent of information apps can quietly gather. The project also highlights that nothing leaves your device unless you explicitly export it. This matters because it reveals the scope of passive and advanced data collection on iOS, and fuels debate about platform-level privacy controls.
Discussion: Mixed — HN commenters are divided: many express frustration over iOS's lack of network permission controls and the ease of fingerprinting, while some defend current design choices as user-friendly. There's admiration for Loupe's transparency, but cynicism about the business incentives behind weak privacy defaults. Comparisons to Android and custom ROMs highlight alternative approaches. (network permission controls, device fingerprinting, privacy by design)
0:00 / 1:06hardware Hyundai Takes Full Control of Boston Dynamics, Aiming for Factory Robots by 2028
Hyundai has acquired SoftBank's remaining 9.65% stake in Boston Dynamics for $325 million, making the robotics pioneer a wholly owned Hyundai subsidiary. This move finalizes a process started in 2021, positioning Hyundai to deploy Boston Dynamics' Atlas humanoid robots in its own EV factory by 2028. The acquisition signals Hyundai's commitment to integrating advanced robotics into manufacturing, while SoftBank shifts its focus to AI infrastructure. The deal underscores growing competition in humanoid robotics, with rivals like Tesla and Figure AI also targeting industrial automation.
Discussion: Mixed — Discussion is split between those seeing the move as a strategic leap for Hyundai and skeptics questioning the utility of humanoid robots versus specialized machines. Some see Hyundai's full ownership as a smart way to 'dogfood' robotics in-house, while others question SoftBank's timing and the real-world viability of humanoid forms in factories. (Ownership transition and deal structure, Strategic use of humanoid robots in manufacturing, Comparison to Tesla and other robotics efforts)
0:00 / 0:56hardware Why Your Screen Can't Show You All the Colors You See
A widely-read blog post explains why digital screens can't reproduce many real-world colors—especially vivid cyans and saturated reds—due to the limitations of standard color gamuts like sRGB. The article details how human vision works, the historical choices behind display technology, and why even modern screens and lighting often fail to capture the full spectrum visible to our eyes. This matters for everything from digital art to accurate scientific imaging, and highlights the gap between digital and real-world color experiences.
Discussion: Positive — The discussion is enthusiastic and appreciative, with many commenters sharing personal experiences and technical insights about color reproduction. While some point out additional limitations and technical nuances, the overall tone is engaged and constructive, with praise for the article's clarity. (Limitations of digital color reproduction, Technical discussion of color spaces and hardware, Personal anecdotes about color in art and technology)
0:00 / 0:48startups StartupWiki Aims to Be the Free, Open Crunchbase Alternative
StartupWiki has launched as a free, accountless database for discovering and researching startups, aiming to offer a Wikipedia-like experience without paywalls or clutter. The project is in early stages, with features like startup profiles, search, categorization, and a public API in progress. Hacker News users appreciate the concept but raise concerns about data accuracy, verification, licensing, and the need for open sourcing and community trust. The founder is actively responding to feedback and iterating on the product.
Discussion: Mixed — The community is enthusiastic about the idea but skeptical about data reliability, verification methods, and long-term openness. There is significant constructive feedback and encouragement, but also concerns about accuracy and trust. (data accuracy and reliability, open source and licensing, community trust)
A web agency, Qontour, created a slick website mimicking John Koenig's 'The Dictionary of Obscure Sorrows,' copying the entire text of his book without permission and replacing original illustrations with AI-generated images. The site also uses AI to generate new 'sorrows' and invites user submissions, all while acknowledging it doesn't own the rights. Koenig confirmed he had no involvement. The incident highlights growing concerns over copyright infringement, especially as AI tools make plagiarism easier and enforcement against such acts remains difficult.
Discussion: Negative — Commenters are largely outraged by the blatant plagiarism, frustrated by the ineffectiveness of DMCA takedowns for individuals, and worried about the ease with which AI enables and scales such theft. There is also debate over copyright, free software, and the power imbalance favoring large corporations over creators. (copyright infringement, ineffective DMCA enforcement, AI-enabled plagiarism)
0:00 / 0:14general Google Reports 50% of Users Now Access Services via IPv6
Google has announced that, for the first time, 50% of its users are accessing its services over IPv6, marking a major milestone in the long-running transition from IPv4. However, adoption is uneven globally, with APNIC Labs measuring worldwide IPv6 capability at 42% and significant variation by country and ISP. The shift to IPv6 is driven by mobile networks and newer entrants, while established ISPs in some regions lag due to legacy investments and limited customer demand. The transition remains complex, with technical, economic, and incentive barriers still slowing universal adoption.
Discussion: Mixed — Commenters are encouraged by the milestone but remain frustrated with slow ISP adoption, uneven regional support, and technical barriers. Many share personal experiences with ISPs not enabling IPv6, while others note improvements in some countries and sectors. There is optimism about IPv6's potential, but skepticism about incentives and the pace of change. (ISP reluctance and incentives, Regional adoption disparities, Technical hurdles for end users)
0:00 / 0:16general Beyond All Reason: A Free, Ambitious RTS Inspired by Total Annihilation
Beyond All Reason (BAR) is a free, open-source real-time strategy game inspired by the classic Total Annihilation. It features fully simulated projectiles, terrain deformation, and massive battles with thousands of units. While the technical achievement and nostalgic appeal are widely praised, many Hacker News users discuss issues with community toxicity and the challenges of online multiplayer culture.
Discussion: Mixed — Commenters are enthusiastic about the game's technical achievements and nostalgic value, but many express frustration with toxic behavior in multiplayer lobbies and the impact of competitive 'meta' culture. Some users share positive experiences in smaller or co-op games, while others lament the shift in gaming communities. (Nostalgia for classic RTS games, Technical innovation in gameplay, Toxicity and competitiveness in online communities)
0:00 / 0:19general Finland's Libraries Lend More Than Books—From Sewing Machines to Democracy
Finland's public libraries, like Helsinki's Oodi, are redefining their role by lending not just books but also items like sewing machines, musical instruments, and even sports equipment. This approach, rooted in the country's pragmatic culture and supported by robust public funding, positions libraries as vital community hubs that promote social inclusion, democracy, and access to resources. The model contrasts sharply with trends in countries like the UK and US, where library budgets have faced cuts. The story has sparked discussion about similar 'Libraries of Things' initiatives worldwide and the broader value of libraries as third spaces and makerspaces.
Discussion: Positive — The overall sentiment is enthusiastic and supportive of the expanded role of libraries, with many commenters sharing personal experiences of similar services in their own communities and expressing appreciation for the concept. Some practical concerns are raised about maintenance and logistics, but these are generally constructive. (Community building, Access to resources, Innovation in public services)
0:00 / 0:15general Windows 11’s New Media Player Bloats RAM and Paywalls Key Codecs
Microsoft’s new Media Player for Windows 11 uses about 3.5 times more RAM than its predecessor and now charges users for popular video codecs like HEVC (H.265) and removes built-in AC-3 (Dolby Digital) support. The app is slower to start and replaces both Groove Music and the classic Windows Media Player as the default. While Microsoft cites codec licensing costs, users and developers are frustrated by the increased bloat and the paywalling of essential features, with many recommending third-party players like VLC that bundle codecs for free. This shift highlights ongoing tensions between user experience, software development convenience, and licensing costs.
Discussion: Negative — The HN discussion is largely negative, criticizing Microsoft for increased bloat, slower performance, and paywalling essential codecs. Many see this as part of a broader decline in native software quality and user experience on Windows. (Software bloat and inefficiency, Codec licensing and paywalls, Declining native app quality)
0:00 / 0:15general SMPTE Opens Its Standards Library to All
The Society of Motion Picture and Television Engineers (SMPTE) has made its entire catalog of standards, recommended practices, and engineering guidelines freely available online. This move eliminates paywalls for critical media technology specifications, aiming to boost interoperability and innovation across the industry. The decision is part of SMPTE's broader modernization efforts, including adopting open workflows and structured publishing. Industry support from major media and tech companies underpins this shift, which is expected to benefit developers, educators, and emerging markets worldwide.
Discussion: Positive — The Hacker News community broadly celebrates SMPTE's decision, viewing it as a long-overdue step toward open, accessible standards. Many see this as a model for other standards bodies, though some discuss the financial realities and historical reasons for paywalled standards. The conversation also highlights frustration with other organizations that still restrict access, such as IEEE and ISO. (Open access to standards, Barriers to entry and innovation, Historical context of paywalled standards)
0:00 / 0:31general TownSquare Adds Real-Time Presence to Any Website—But Moderation Looms Large
TownSquare is a lightweight widget that lets visitors see and interact with each other on any website, bringing a playful, real-time presence layer without requiring accounts or algorithms. The project quickly gained attention for its simplicity and fun, but the live demo was immediately beset by offensive spam, highlighting the persistent challenge of moderating open, anonymous spaces. While many users love the concept, concerns about abuse and the need for effective moderation—possibly via LLMs or client-side filtering—dominated the discussion. The experience underscores both the appeal and the pitfalls of adding social layers to the web.
Discussion: Mixed — Commenters are enthusiastic about the concept but deeply concerned about abuse and moderation challenges. While some celebrate the fun and nostalgia, many see the spam problem as a dealbreaker unless better controls are implemented. (Moderation challenges, Nostalgia for early web, Abuse of open platforms)