The newest HackerNews stories, desensationalized.

A developer describes bringing Linux to Apple’s M4 Mac mini despite new barriers introduced by Secure Page Table Monitor, which complicated m1n1-based reverse engineering and hypervisor support. Early progress required disabling strict boot security, skipping initialization of locked GXF and RVBAR registers, and using serial debugging. A minimal device tree and early assembly prints revealed that Linux lacked identity mappings for UART MMIO after enabling the MMU, while a virtualization register write also caused crashes. Adding the mappings, correcting stdout-path, and avoiding that register write allowed Linux to reach a shell. Enabling secondary cores exposed an M4 flaw in which WFI and WFIT could erase architectural registers, contrary to the ARM64 specification. Replacing those instructions with NOPs enabled multicore booting, leading to upstream Linux support for a boot option that disables WFI idle and corresponding m1n1 changes. Current systems can boot M4, M4 Pro, M4 Max, and M5 machines with all cores, while peripheral reverse engineering and hypervisor support remain ongoing.

A federal judge issued a preliminary injunction against Utah’s SB 73, which would require adult websites to identify visitors’ physical locations, block Utah users, or verify everyone’s age, while also restricting VPN-related instructions. The court found that the law likely violates constitutional limits on state laws with substantial effects outside Utah and effectively imposes strict liability because perfect geolocation is not technically possible. VPNs conceal users’ originating IP addresses by routing traffic through intermediary servers, making locations such as Salt Lake City, Seattle, and Shanghai difficult to distinguish reliably. The proposed enforcement rules would have required detection of geolocation obfuscation using unreliable signals such as latency and device time zones, creating surveillance, misclassification, access-blocking, and privacy risks. The ruling pauses enforcement while litigation continues, although it does not challenge the separate restriction on VPN information. EFF argues that similar laws will encourage broader tracking and motivate VPN providers and users to adapt, while Utah lawmakers may revise the law.

Cloudflare is launching a paid, closed-beta OHTTP Gateway that lets applications receive privacy-preserving HTTP traffic while keeping client identifiers separate from request contents. OHTTP uses two independently operated hops: a relay that sees client metadata and forwards ciphertext, and a gateway that decrypts requests and encrypts responses without seeing client identifiers. Cloudflare’s existing OHTTP Relay, formerly Privacy Gateway, remains suitable when customers host their own gateway, while the new Gateway fits applications hosted on Cloudflare or receiving traffic from third-party relays. The managed service runs across Cloudflare’s edge, supports standard and chunked OHTTP, manages HPKE keys, authenticates relays through Cloudflare Access, limits forwarding to the customer’s zone, and rejects traffic from Cloudflare-operated relays or proxied hosts to preserve separation of trust. Customers must supply a relay and avoid placing identifying data in request bodies.

Former Pittsburgh Steelers head coach Mike Tomlin has spent about 12 years building a large modern city in Minecraft, beginning when his children were young. The project was created in the game’s creative mode, primarily with a controller and relatively slow sensitivity settings, making it especially time-intensive. Viewers found the city technically modest by experienced Minecraft-building standards but valued its personal character, sustained effort, and connection to family life. Tomlin’s long NFL tenure and winning record made the hobby seem especially surprising, while one humorous comparison linked his city-building period with a decline in playoff success. The story prompted broader reflections on play, creative work without a commercial objective, automation and AI-generated construction, and the possibility that greater economic security could give more people time for meaningful hobbies.

Pass Designer is a macOS 27 beta tool for creating Apple Wallet passes for businesses and organizations. It provides Apple and custom templates, supports externally created logos and other images, and previews passes in real time using the same rendering as iOS and watchOS. Designers can edit colors, layouts, standard fields, and semantic tags for boarding passes and event tickets. Semantic data can support Siri Suggestions, Calendar integration, and Maps directions, while the tool can generate backward-compatible structures for systems without semantic-tag support. It also validates passes during editing and reports missing required values or invalid definitions. Download requires an Apple Account, free registration if prompted, and acceptance of the Apple Developer Agreement.

A 12-year sequence of telescope observations shows four exoplanets, each more massive than Jupiter, orbiting a star about 133 light-years away. The imagery combines observations from multiple telescopes and wavelengths, while another animation uses only Keck data from one instrument at 3.5 microns. Viewers emphasize that the result is not continuous video: it contains about ten real images with interpolated frames. The discussion explains that direct imaging is difficult because the planets are faint and close to a much brighter star, and asks about artifacts, observing frequency, stellar variability, hidden smaller planets, and improved processing. Commenters also highlight future capabilities from the Nancy Grace Roman Space Telescope’s coronagraph and the planned Habitable Worlds Observatory, which could image reflected light from Jupiter-like and eventually Earth-like planets. The sequence is valued both for its scientific significance and for making exoplanet motion visually accessible.

Aging may involve not only accumulated cellular damage but also erosion of epigenetic programs that preserve specialized cell identity. A three-tier regulatory system responds to acute stress, manages longer cell-state transitions, and maintains durable chromatin architecture; chronic inflammation can overwhelm these systems and impair PRC2-mediated constraints. As Waddington’s epigenetic landscape becomes less stable, differentiated cells may drift toward a fibroblast-like mesenchymal state, promoting extracellular-matrix deposition, fibrosis, inflammation, disease progression, and possibly cancer. This drift has been observed across 46 tissue types and is associated with conditions including atherosclerosis, macular degeneration, and Alzheimer’s disease. Caloric restriction may help preserve epigenetic architecture, while partial Yamanaka-factor reprogramming has restored cell identity in cells and animals but carries substantial safety risks. Lithium-related pathways are another experimental possibility. The model may clarify what epigenetic clocks measure, but its relationship to accumulated damage, programmed aging, species differences, and human rejuvenation remains unresolved.

Stratego remained difficult for AI because it combines 40 hidden pieces, more than a decillion possible setups, games lasting up to 2,000 moves, and bluffing under uncertainty. A research team from Carnegie Mellon, MIT, NYU, and Stanford developed Ataraxos, which defeated elite player Pim Niemeijer 15–1 with four draws after training on 16 GPUs costing a few thousand dollars. The approach reportedly used about 34 times fewer games than DeepMind’s 2022 DeepNash while achieving stronger play, suggesting that rapidly learning opponents’ behavior can compensate for the impossibility of exhaustively searching unknown states. The result also points to broader applications in imperfect-information games such as Hanabi, while games with frequently changing rules may remain harder because new content can invalidate training data.

Meta is offering U.S. subscribers with active Muse memberships a free Muse Home Link, limited to one per subscriber and shipping in October while supplies last. The device connects Muse to compatible home-network devices and locally hosted HTTP APIs, enabling actions such as switching lights, controlling televisions, and sending documents to printers. Community-built integrations and the Muse Gadget SDK are aimed at personal experimentation, including beginner-friendly, vibe-coded projects and microcontroller hardware such as ESP32 boards. The SDK requires device tokens, supports only a limited number of paired devices, and may change or stop working without notice, so it is not presented as a supported developer platform. The initiative could help Meta bootstrap a broader hardware and agent ecosystem by encouraging enthusiasts to connect physical devices to Muse.

Researchers criticize four Harvard studies linking proximity to nuclear power plants with cancer or total mortality, arguing that geographic proximity is not a measured exposure and that the analyses lack a plausible radiation pathway. They replicated the approach imperfectly and applied it to landmarks including Costco, universities, sports venues, capitals, and Superfund sites, obtaining large attributable-death estimates even for places with no credible mechanism. They argue the method may generate positive associations from its design, covariates, ecological data, and broad geographic ranges, including when outcomes are randomized, so its precise estimates may be statistical artifacts rather than causal effects. The criticism emphasizes inadequate controls, absent placebo tests, and failure to test alternative landmarks before interpreting nuclear results, while acknowledging that the original studies described associations rather than proof and that unexpected findings warrant technically transparent investigation.

DwarfStar 4 is a MIT-licensed C inference engine for high-memory Mac, CUDA, and ROCm systems, focused on selected DeepSeek V4 and V4.1 Flash, GLM 5.x, and Qwen3.8 Flash Next models. It uses asymmetric quantization to compress routed experts while preserving shared critical paths, project-specific GGUF files, and SSD-backed prompt-prefix caching to make large mixture-of-experts models practical on consumer hardware. The stack includes a CLI, OpenAI- and Anthropic-compatible server APIs, and a persistent coding agent, with text and vision support on some backends. Reported reference results include 34.4 tokens per second generation on an M5 Max with 128 GB and Q2 weights, while long contexts reduce throughput. The project is specialized rather than a generic GGUF runner, and installation requires matching its models and hardware guidance.

FLUX 3 Image is Black Forest Labs’ multimodal image-generation and editing model, focused on controllable composition through a 0–1000 canvas and editable bounding boxes. Users can describe each element in a JSON table, combine it with a global scene prompt, and generate images with defined spatial relationships, including text, crowds, and complex layouts. Individual regions can later be replaced, moved, or re-described while untouched areas remain consistent. The model supports multi-reference composition, pixel-level edits, and native 2K and 4K rendering. It is available under a commercial-weights license for scaled deployment, while an open release is anticipated but not yet available. The presentation emphasizes an agent-oriented workflow and a less cumbersome interface than some competing structured prompting systems.

Apple says macOS Full Disk Access bypasses normal privacy protections so backup software can function, but some applications may consequently read files, mail, messages, browsing history, and communications data without users fully understanding the consequences. It plans stronger controls requiring explicit consent, especially as increasingly autonomous AI agents make broad access riskier. The change aims to improve informed privacy decisions, but the proposal gives few implementation details and may affect legacy software, development workflows, command-line tools, screenshots, and other utilities. The central technical request is more granular, auditable permissions, such as read-only access, file-size-only queries, selected folders, and clear revocation controls.

Researchers from the University of British Columbia identified seven sea spider species in the Salish Sea, including five known species and two previously undescribed species, the first regional discoveries in about a century. Specimens collected from shorelines to roughly 60 feet deep were classified using scanning electron microscopy and DNA analysis. Sea spiders are ancient marine arachnids related to scorpions, spiders, and horseshoe crabs; they have jointed exoskeletons, piercing or grasping mouthparts, breathe through their skin, and can carry eggs on specialized male appendages. Callipallene pilosuspedes is about one centimeter long and has hairy feet, a triangular, sensory-hair-covered mouth, and males that generally lack claws on their egg-carrying appendages. Tanystylum kiixin has short, stiff legs that collect debris, which supports tiny single-celled parasites and creates a miniature ecosystem on its body. Some shallow-water species also parasitize larger organisms by using a proboscis to extract fluids.

ChatGPT Sites lets users describe an idea, attach files or visual references, build a website, app, or game through conversation, and publish it without separately configuring hosting or domains. Sites can be private, workspace-restricted, selectively shared, or public, with optional ChatGPT sign-in for personalized experiences. The positioning emphasizes rapid prototyping, broad use cases, and ongoing editing. The comments suggest that the service can produce working prototypes quickly and may reduce deployment friction, but they also report slow performance, inconsistent quality, broken demonstrations, shallow visual effects, unclear APIs for dynamic data, and uncertainty about version control, resource limits, and long-term support. Commenters debate whether this threatens web developers, marketing teams, and app-building companies, or instead increases demand because generated sites often require human judgment, iteration, and technical integration. Several expect AI companies, device makers, and hosting providers to compete for an app platform, while others question whether hosted site builders provide a durable competitive advantage.

AI models created oil paintings by programming a simulated oil medium on linen, including bristle behavior, wet paint, drying, and layered glazes, without using image generators or viewing Friedrich’s paintings. Some models wrote complete programs, while others worked passage by passage at a virtual easel using image looks, journals, and studio notes. Across rounds, models repeatedly selected dusk scenes, jugs with lemons, bare trees, Baltic shores, and even identical titles without seeing one another’s work. The project also exposed limitations: one model’s image-context bug caused it to inspect stale canvases, another noticed background evaluation activity, and several paintings remained unfinished because of usage limits. A prompting error led one painter to avoid Friedrich motifs after treating earlier notes as rules.

A web tool predicts the locations of directly imaged exoplanets and brown-dwarf companions. Its creators are Jason Wang, Matas Kulikauskas, and Sarah Blunt. Researchers can request additional orbital data from Wang and should cite the tool’s ASCL record when using it in research.

NVIDIA’s updated Zorah scene demonstrates clustered ray tracing and Nanite-like hierarchical LOD for a 1.6-billion-triangle path-traced environment. Meshoptimizer improvements reduce geometry storage through fast vertex, index, and meshlet compression, with v1 compression cutting source geometry to about 31.3 GB and meshlet encoding reducing processed cluster data from 54.9 GB to 47.0 GB. These codecs decode quickly enough that compression adds little processing or streaming cost and can be layered with general-purpose compressors. Processing improvements include spatially coherent cluster hierarchies, attribute-aware simplification, more reliable permissive mode, reduced allocation traffic, and clamped attribute error, which lowered tested geometry and acceleration-structure memory needs substantially. Foliage required boundary-area preservation to avoid distant hedges eroding, though the technique can create artifacts and remains opt-in. The complete scene processes in roughly four minutes on 16 CPU cores and runs on a 12 GB RTX 5070, but clustered LOD still carries storage, runtime, memory, complexity, and visual-popping tradeoffs.

OKI Circuit Technology developed a 124-layer printed circuit board for wafer-inspection equipment used with next-generation high-bandwidth memory in AI semiconductors, increasing layer count about 15% over conventional 108-layer designs while maintaining a 7.6-mm thickness. The design addresses rising signal counts and chip densities caused by semiconductor miniaturization and the need for faster, higher-frequency, higher-density GPU-to-memory data transfer. OKI attributes the advance to ultra-thin materials, specialized tools and handling methods, and a proprietary automated transport system for those materials. The company planned to establish mass-production technology at its Joetsu Plant by October 2025 and presented the technology at PCB East 2025. OKI said the underlying PCB capabilities also support anticipated demand in aerospace, defense, robotics, and next-generation communications.

Change coupling describes code that repeatedly changes together, even when static dependencies, file locations, or architectural boundaries do not reveal the relationship. It exposes hidden technical, business, regulatory, temporal, and organizational dependencies that can scatter small changes across many files, increase cognitive load, require cross-team coordination, and eventually make critical code difficult to modify. A physical-versus-change-coupling matrix helps identify code that should potentially be consolidated or split, but it is a sense-making tool rather than a metric to minimize. The payment-system example shows why merging change-coupled modules can create a monolith, while dependency inversion or targeted restructuring may isolate shared rules and machinery. Events, microservices, interfaces, domain boundaries, and architecture patterns do not eliminate logical coupling when two parts still need to change together. Teams should analyze historical co-change patterns, prioritize the changes that consume the most time or create the most risk, choose context-specific structural improvements, and review the patterns regularly because business strategy, team organization, and system needs evolve.

In 1171, Byzantine forces seized more than 20,000 Venetians and their property amid escalating commercial and political tensions. Venice responded by raising an involuntary loan, promising 5% annual interest and making the claims transferable, allowing citizens to sell or collateralize them. The resulting fleet initially attacked Byzantine territory but stalled during negotiations, suffered a devastating plague outbreak, and returned defeated; the doge was then killed by an angry mob. Venice later maintained interest payments despite being unable to repay the principal, supported them with Rialto tax revenue, and kept transparent records under institutions increasingly overseen by the Great Council. Tradable claims became a durable form of public debt, linking citizen wealth to state solvency and helping create the Rialto’s early bond market. The crisis also encouraged institutional reforms, including a commission-based doge selection process and a stronger proto-parliament. Similar debt instruments were later adopted by other Italian city-states.

LDraw Nova is a Dockerized web app that lets AI agents design LEGO models from prompts and output LDraw source, 3D views, VR experiences, images, and editable Blender .glb files with metadata. Agents read specialized documentation, search parts and example constructions, create JSON plans and Python generators, render intermediate results, inspect geometry, detect collisions and gaps, and iteratively revise models before producing .mpd files. The project argues that agents handle Python abstractions and reusable building primitives better than direct LDraw geometry and rotation math. Semantic part and example searches use the author's jev-rerank tool when a TypeSafe API key is configured, with full-text search as fallback. The software is released at version 0.6.0 with companion Docker configuration, but large correct models require expensive high-end models and generation remains slow. VR performance, low-end-model support, Technic and spaceship workflows, human and animal models, manual reconstruction, and fine-grained inspection still need improvement.

The author describes a static-blog workflow centered on writing in Emacs with Org-mode, while using D2 and Babel for executable diagrams and content, Pandoc to convert Org files into cited GitHub-Flavored Markdown, Gleam and Blogatto to generate pages and RSS, Lustre for typed HTML layouts, and Nix with devenv to reproduce build dependencies. Org files remain the only source of truth, while Lua filters handle verses, bibliographies, and image paths. The author values Org-mode because its interactive document model combines prose, source code, data, citations, and computation in one environment, arguing that this coherence can outweigh its dependence on Emacs. Pandoc is used as a deliberately simple boundary because it already supports the required conversion and citation features. The author acknowledges that the stack is elaborate for a static site and that Org-Publish could replace parts of it, but prefers the quality-of-life and type-checking benefits of the chosen tools.

The listing states that a successful bidder must pay the final bid plus the buyer’s premium, applicable taxes, any artist resale royalty, and shipping-related expenses, and notes that a specialist may provide or update the condition report. The comments identify the lot as a framed printed IP/ICMP packet sent by carrier pigeon under RFC 1149, and question whether the high price reflects the frame rather than a scarce artifact. Others joke about preserving old USB drives and using RFC 1149 in email signatures, while commenters debate whether 2001 should be characterized as the early Internet and question the claim that the period was especially collaborative. Suggestions include adding a stuffed pigeon and using a prediction market for the sale price.

A month-long attempt to use GLM 5.3 Flash exclusively achieved only about 50% adherence: 1B of 2B tokens went to other models. The target model cost about $68 and was estimated at 4 kWh and 365 grams of carbon, but a vibe-coded Wagtail MCP prototype consumed 450M tokens, $150, and 5 kWh after an unsuitable model and agentic workflow were chosen. Infrastructure capacity problems also forced switches to DeepSeek V4.1 Flash and Qwen 3.8 Flash. The experiment still produced a working demo, benchmark data, and lessons about model selection, budgeting experimentation, prompt design, multi-agent roles, and measuring cost and energy alongside outcomes. The author concludes that one or two inexpensive flash-tier models are practical for routine development, while experimentation should be separately budgeted and measured.