Hacker News

The Front Page

Vol. I November 12, 2026 Price: A cup of coffee

19h, 1265p, 141 comments

Tim Peters has died

Guido van Rossum and the Python community are mourning the passing of Tim Peters, who died on July 15, 2026. A legendary figure in the development of the Python programming language, Peters was the author of the Zen of Python (PEP 20) and the creator of the Timsort algorithm, which was later adopted by Java and Android. Colleagues and fellow developers remember him as a brilliant mathematician and compiler expert, often referred to by the nickname “Timbot” for his patient and humorous guidance on mailing lists. Beyond his technical contributions to core features like garbage collection, generators, and numerical algorithms, Peters is celebrated for his mentorship, his lack of pretension, and his ability to foster a welcoming, inclusive culture within the Python Software Foundation.

17h, 456p, 143 comments

Gemini Decisions API

Google has introduced the Gemini Decisions API, a new tool powered by the Gemini Decision 1 model designed for fast, probabilistic, and multimodal decision-making. The API allows users to evaluate up to 1,000 questions—including yes/no, multiple choice, and rubric scoring—in a single pass across various inputs such as text, images, audio, video, and PDFs. By scoring candidate options directly rather than generating text token-by-token, the model achieves low end-to-end latency of 150-250 ms and is significantly more cost-effective, priced at $0.10 per 1M input tokens with zero charges for output tokens. Currently available through Google AI Studio, the API is being utilized for tasks like dynamic model routing in Gemini Notebook and safety monitoring in Google Search.

1d, 533p, 540 comments

Apple has new Mac app coming to make running local AI models easy

Leaker Aaron Perris has revealed evidence of a new upcoming Mac application called Harbor, designed to transform a Mac into a private AI inference server. Developed under Apple's Claris brand and currently in private preview, the app allows users to run open-weight models locally on Apple silicon, ensuring data remains on the hardware without cloud involvement. A standout feature of Harbor is its ability to cluster multiple Thunderbolt 5-capable Macs to pool unified memory, enabling the execution of large-scale models that exceed the capacity of a single machine. The application aims to simplify the process by offering automated device detection and built-in model management, removing the need for command-line expertise or complex infrastructure setup.

12h, 21p, 1 comment

Porting my Electron app to Rust: 675 MB to 91 MB

Developer Mohak Bajaj has announced the release of Jot 2.0, a significant overhaul of his Mac capture and dictation application from an Electron-based framework to Rust using the Rune GPU UI framework. This architectural shift has drastically reduced the application's disk footprint from 675 MB to just 91 MB and streamlined the download size from 302 MB to 44 MB. By moving away from Electron's multi-process IPC model to a single-process structure where view and state are unified, the app now offers improved performance and cross-platform compatibility for Windows and Linux, in addition to its primary macOS support. The update maintains full parity with previous versions, allowing users to update in place while sharing the same Markdown-based vault and settings. Furthermore, the port includes enhanced dictation features, such as improved silence detection and faster model warming, to provide a more seamless user experience.

19h, 118p, 109 comments

AI, activity, and outcomes in software engineering

In a seminar held at UC Irvine on October 30, 2026, Christian Bird of the Dryden Technology Group addressed the paradox of why software shipping speeds have not significantly increased despite the integration of AI. Titled "Why aren’t we shipping faster now? AI, activity, and outcomes in software engineering," the presentation explored the activity-outcome gap, suggesting that while AI tools may increase individual coding activity, they do not necessarily translate into faster delivery of completed software. Bird supported his discussion with several 2026 research papers examining the productivity effects of different generations of AI coding tools, the boundaries developers set regarding AI autonomy, and the factors driving the adoption of Generative AI within individual and team environments.

1d, 286p, 217 comments

Claude Whiteboards, Claude Sheets, and Claude Data

On November 10, 2026, Anthropic announced the launch of Claude Whiteboards and Claude Sheets in beta, alongside the transition of Claude Dashboards to general availability as Claude Data. These new features, available on all paid plans, aim to allow teams to brainstorm, build spreadsheets, and analyze data directly within the Claude interface. Claude Whiteboards facilitates real-time collaboration through visual boards and timelines, while Claude Sheets enables the AI to generate and edit spreadsheets with formulas and formatting. Claude Data expands on earlier dashboard capabilities by offering certified queries, live data sharing, and data storytelling across various data warehouses. Additionally, Anthropic introduced interactive connectors, allowing users to work with integrated partner applications like Adobe, Atlassian, and Notion directly within the platform.

1d, 924p, 204 comments

.name Grandfathered

In a major victory for domain owners, Verisign has retracted its plan to terminate 22,000 .name domain names and email redirects. While new registrations for 3rd-level .name domains will be closed, all existing domains and email redirects will be grandfathered and remain active indefinitely. This outcome follows an extensive campaign involving legal experts and numerous advocates who challenged the decision made by ICANN.

5h, 29p, 9 comments

Why Every AI Model Will Ship as a Torrent

In an article titled "Why Every AI Model Will Ship as a Torrent," the platform Pirate Face argues that BitTorrent protocols are the ideal method for distributing large-scale AI models like DeepSeek-V4.1-Flash-1017. Unlike movies, which have various resolutions and versions, AI models consist of fixed, immutable files where every user requires the exact same bytes, making them perfect for a single, unified torrent swarm. The piece explains how the v1 (BEP 3) and v2 (BEP 52) torrent formats use cryptographic hashes to ensure content integrity, and highlights the role of trackers and Distributed Hash Tables (DHT) in helping users find peers without central servers. Pirate Face aims to serve as a front-page directory for the Model DHT, providing verified magnet links and real-time swarm data to support the availability of sovereign AI models.

2d, 168p, 120 comments

M John Harrison becomes first science fiction author to win the Booker

In a historic achievement, 81-year-old author M John Harrison has become the first writer from the bona fide science fiction genre to win the Booker Prize for his postapocalyptic novel, The End of Everything. Set on the near-future Kent coast, the 198-page novel follows characters attempting to navigate a world altered by mysterious, shape-changing beings known as the iGhetti. Harrison’s win makes him the oldest recipient in the award's history, surpassing Margaret Atwood. Judges described the work as a profound exploration of the "failure to comprehend" disaster, utilizing speculative elements to mirror contemporary issues such as political instability and digital uncertainty. The victory marks a significant moment of mainstream literary recognition for a writer who has spent decades working at the intersection of speculative and literary fiction.

19h, 215p, 38 comments

Faster Retrieval for Better, More Efficient RAG

Researchers from Apple Machine Learning Research have introduced a new approach to optimize Maximum Inner Product Search (MIPS), a core component in recommendation systems and Retrieval-Augmented Generation (RAG). Presented at ICML 2026, the work, titled "Amortizing Maximum Inner Product Search with Learned Support Functions," moves away from traditional query-agnostic methods that rely on geometric clustering. Instead, the researchers use small, pre-trained neural networks to learn how to route queries toward their best matches based on specific query distributions. The study highlights two specialized architectures: SupportNet, which approximates support functions to improve cluster routing, and KeyNet, which acts as a query rewriter to predict the best-matching key vector. Testing on the Natural Questions dataset demonstrated that these learned models can achieve significantly higher recall at lower latencies compared to standard centroid-based routing and existing Inverted File Index (IVF) systems, effectively trading upfront training costs for faster, more accurate retrieval during inference.

23h, 204p, 121 comments

We need more goat sacrifices

In his essay We need more goat sacrifices, Adam Mastroianni argues that modern social policies and philanthropic interventions frequently fail because of a fundamental "illusion of control." Using various examples—from ineffective School Improvement Grants and "One Laptop Per Child" programs to failed tuition waivers at New York University—Mastroianni highlights how well-intentioned efforts often yield zero net impact. He references sociologist Peter Rossi’s satirical "Iron Laws of Evaluation," which suggest that large social programs often have null effects, and critiques the psychological fallacy that humans can exert significant influence over complex systems. Mastroianni concludes that rather than succumbing to cynicism or resignation, society must acknowledge that the social sciences are still in an "alchemical stage," requiring more rigorous, fundamental research to move beyond mere guesswork toward effective interventions.

14h, 34p, 1 comment

Optimizing Reference Deletions in Git

An optimization for Git's files backend has been identified and merged into the Git master branch, set to debut in version v2.98. The discovery occurred while Karthik Nayak was benchmarking a Rust-based implementation of the backend against the existing C code. The investigation revealed that the standard C implementation was inefficiently rebuilding unmodified references using fprintf() and oid_to_hex() during deletion processes, rather than writing them verbatim. By implementing a patch that uses fwrite() to copy data directly, reference deletion speed improved by approximately 18%, bringing the performance of the legacy backend in line with newer Rust-based prototypes.

1d, 273p, 69 comments

A great old-timey game-programming hack (2013)

In this retrospective essay, Tom Moertel recounts a complex 8-bit programming challenge from his college years involving a side-scrolling game on a 6809 microprocessor-based PC. To overcome unplayable frame rates caused by slow background tile rendering, Moertel and a friend developed a series of increasingly "crazy" optimizations. Their ultimate solution involved a 70 percent speed-up achieved by commandeering both the S and U stack registers to perform bulk memory copies via PULS and PSHU instructions. This method initially caused visual corruption because it drew tiles upside down and in swapped 7-byte "septets," which they corrected through a one-time pre-processing step of the tile data. When hardware interrupts for sound effects threatened to corrupt reference tiles by overwriting the system stack, they implemented a final "perfect hack": swapping the stack pointers to target the screen buffer instead and reversing the drawing order (bottom-to-top, right-to-left) so that any transient corruption would be automatically repaired by the subsequent tile draw before the frame was displayed.

3d, 1255p, 167 comments

Foldable Flight: The Best Paper Airplanes

Foldable Flight offers an extensive collection of paper airplane designs, ranging from very easy to expert folding difficulties. The platform features world-record-holding models like the Unite, which holds the record for time aloft, and the Firefly 5, a distance prototype capable of flying over 300 feet. Users can explore various plane types, including darts, gliders, hybrids, jets, and stunt planes, designed by experts such as Kyle Boyer, Michael LaFosse, and Evan Brus. The site provides detailed performance metrics for each model, including speed, glide, and accuracy ratings, alongside tutorials for iconic designs like the Arrowhead and the F-14 Tomcat.

18h, 40p, 8 comments

The tiny transmitter tracking monarch butterflies' migration, Canada to Mexico

Conservationists are utilizing ultra-lightweight BlūMorpho transmitters to track the 4,000-km migration of the "super generation" of monarch butterflies from Canada to Mexico. Developed by Cellular Tracking Technologies and used in the multi-national Project Monarch, these solar-powered devices weigh only 60 milligrams and allow scientists to map precise migratory routes via smartphone signals. This data is critical for identifying "habitat gaps" and essential "gas stations" where butterflies need nectar to survive. By pinpointing where populations decline, organizations like the Assiniboine Park Zoo and the Kootenay Native Plant Society aim to guide the creation of "monarch waystations"—community-led plots of native milkweed and nectar-producing plants designed to support the endangered species throughout its transcontinental journey.

15h, 51p, 12 comments

Tools for exploring the souls of large language models

In his article, Michael OShea introduces the TinyGPT Visualizer, an open-source tool designed to aid in the field of mechanistic interpretability. As researchers increasingly seek to understand the internal processes of Large Language Models (LLMs), OShea developed this tool to function like an "X-ray machine," allowing users to visually explore the numeric matrices and patterns that constitute an AI's "thought" processes. Amidst criticisms that major companies like Anthropic and OpenAI prioritize model performance over safety research, the TinyGPT Visualizer aims to democratize the study of how these models generate prose, code, and other outputs by making the underlying data more accessible to a broader audience. The project is hosted on GitHub at github.com/oshea00/tinygpt-viz.

20h, 79p, 31 comments

Show HN: Klad, a paint app your coding agent can drive (source-available)

Klad is a 2D painting and animation application designed specifically for agent/human collaboration. Unlike standard image generators that produce static pixels, Klad treats every brushstroke, shape, and fill as editable data, allowing users and coding agents (such as Claude Code or Codex) to work simultaneously on the same canvas. While humans use a pen in the editor, agents interact via a command line interface, utilizing the same brushes and layers to execute precise artistic tasks. The software features a full suite of professional tools, including a timeline with onion skins, IK bones, and support for exporting directly to Godot. Currently in an early stage and available via source build, Klad is released under a source-available license that is free for personal use and for companies with annual revenue under 500,000 euros.

1d, 248p, 62 comments

Show HN: Valencia, Spain in Sim City/Isometric View

Tiny València is an isometric pixel-art map of València, designed in the spirit of SimCity and Isometric NYC. The project features a block-by-block recreation of the entire municipal area, spanning from the huerta to the Albufera. To create the seamless landscape, real 3D city data was rendered from a fixed camera and subsequently repainted tile by tile using an image model trained in a specific pixel-art style. The map utilizes data from sources such as OpenStreetMap, Catastro, IGN, and Google Photorealistic 3D Tiles.

22h, 52p, 28 comments

Hooked on Tuna

A collaborative study involving 18 global researchers has revealed that human harvesting of tuna dates back at least 44,000 years, with the earliest evidence found in Jerimalai Cave in East Timor. While ancient fishing practices—ranging from Mediterranean commercial fisheries to Indigenous traditions in North America—involved specialized techniques like harpooning and traps, the scale of extraction was minimal compared to the modern global industry. Today, the tuna market is valued at approximately US$40 billion annually, removing millions of tons from the ocean each year. This massive increase in exploitation has placed species such as the southern bluefin, Pacific bluefin, and bigeye tuna in endangered or vulnerable positions, highlighting a stark contrast between millennia of sustainable harvesting and recent century-scale industrial pressure.

11h, 39p, 11 comments

How to make a CNC plasma cutter with high-frequency start

In a personal project documented by James Stanley, the author details the construction of a DIY CNC plasma cutter designed to overcome electromagnetic interference (EMI) caused by high-frequency (HF) start machines. To prevent the HF arc from disconnecting USB connections, Stanley implemented a solution using an ESP32 board running FluidNC, allowing communication with the motion controller over WiFi. The machine features a structure made of steel box section and 3D-printed ABS-GF parts, utilizing Leadshine easy servos powered by a 12V supply and GT2 belts. While the build successfully allows for automated cutting, Stanley noted challenges such as inconsistent cutting depth, beveling issues, and the necessity of isolating power loads to prevent tripping breakers. He concludes that for HF start machines, minimizing electronic inputs—such as limit switches and USB cables—is crucial to maintaining stable operation.

22h, 102p, 18 comments

Show HN: Godbolt-in-a-floating-window for Neovim (C, C++, Rust)

AssemblyForNeovim is a lightweight Lua script developed by 0xdeadf1sh designed to display inline assembly within a floating window in Neovim. Supporting both x64 and aarch64 architectures, the tool provides assembly views for code written in C, C++, and Rust. Users can trigger the functionality via specific keybindings in normal and visual modes, or through the :Asm command. The plugin requires cc, c++, rustc, or cargo to be available on the system's PATH and can utilize compile_commands.json for compilation details. Within the assembly window, users can adjust optimization levels using the + and - keys, rerun the process with R, or exit with q.

1d, 198p, 48 comments

Why I wrote the BEAM book (2025)

After a decade of managing core systems at Klarna, engineer Erik authored The BEAM Book to provide a deep technical manual for the Erlang/Elixir virtual machine. The journey to publication was marked by years of persistence, including two cancelled publishing deals with O’Reilly and Pragmatic Bookshelf, and numerous technical rewrites. Transitioning the project to a public GitHub repository allowed community contributions to drive the book toward completion. The final work serves as a comprehensive guide for engineers operating large-scale systems, covering critical topics such as schedulers, process management, memory allocation, garbage collection, and performance tuning. The paperback version of The BEAM Book 1.0 is now available via Amazon.

13h, 92p, 6 comments

I Accidentally Became a Millionaire

Security researcher Lyra Rebane detailed a viral incident in which she appeared to become a millionaire after mistakenly gaining edit access to a Google Buganizer issue. While acting as a verifier for a Chromium vulnerability, Rebane used her unauthorized permissions to change a bug bounty field from $1,000 to $1,333,337 as a joke. The stunt went viral on X (formerly Twitter), leading AI-generated news sources and Google's AI Overview to report the fake payout as factual. Rebane used the event to highlight the fragility of modern information channels, noting how AI disinformation machines and flawed Community Notes can rapidly amplify and solidify falsehoods.

14h, 270p, 56 comments

Lock Your GitHub Actions

In a recent blog post, Hugo van Kemenade outlines how to improve GitHub Actions security by utilizing workflow lockfiles. To mitigate supply-chain attacks caused by mutable tags, Kemenade recommends using the gh-actions-lock extension for the GitHub CLI to transition from manual SHA pinning to a more manageable lockfile system. This method generates a .github/workflows/actions.lock file that tracks specific commits for each action. The author highlights that Renovate and Dependabot already support these lockfiles, and suggests validating them in CI using a verification job. Additionally, he recommends using zizmor, a security linter, to ensure workflows adhere to modern security standards, noting that versions from v1.32.0 onwards specifically support lockfiles.

22h, 400p, 76 comments

Hyper-Realistic Sand with WebGL2

Sand is a hyper-realistic sand simulator built with WebGL2 and JavaScript that allows users to interact with granular physics directly in their web browsers. Compatible with phones, tablets, and desktops, the tool enables players to push grains with a finger, tilt their devices to slide the sand, or shake them to scatter the particles. The simulation offers various interactive capabilities, such as pouring, drawing, and support for haptics and sound to enhance the user experience.

7h, 28p, 6 comments

Darling Development Update – Q4 2026

In the Q4 2026 Development Update, Darling founder Luboš Doležel announced a significant shift in the project's direction, headlined by the adoption of a formal AI-assisted contribution policy. To combat the massive scale of Apple's proprietary frameworks and the limited number of macOS specialists, the project will now leverage LLMs to accelerate development, provided that human contributors remain responsible for the code and adhere to strict clean-room implementation rules. A major technical milestone was also reached with the introduction of initial ARM64 support, essential for compatibility with modern Apple silicon hardware. Looking ahead, the project aims to merge various forks—including VibeDarling—back into the main codebase, stabilize AppKit support, and officially support i386, x86-64, and ARM64 architectures.

18h, 67p, 70 comments

The hardest notes to find on the guitar

An analysis of 275,169 data points from Gitori's "Find the Notes" game reveals significant patterns in how guitarists identify notes on the fretboard. According to author Udit M., the most difficult area for players is the upper middle of the neck, specifically the G string from frets 6 to 9 and the B and D strings between frets 7 and 9. The single most challenging spot is the G string at fret 8 (E♭), which is answered incorrectly roughly one-third of the time. Conversely, the edges of the neck, including the low E string, high E string, and the 12th fret, are the easiest to navigate. The data also indicates that accidentals—particularly F♯, B♭, A♭, and E♭—are harder to identify than natural notes, resulting in both increased error rates and slower identification speeds.

1d, 1170p, 438 comments

Smart home devices should work without the internet

In a detailed long-form essay, Spencer Kittleson advocates for a "local-first" philosophy in home automation, arguing that core smart home functions must remain operational even when the internet is unavailable. Drawing from five years of experience managing a multi-user household, Kittleson outlines essential rules to avoid the failures of modern "cloud-dependent" devices, such as smart appliances that become useless when a vendor's server goes down or a company ceases operations. His setup relies on hardware he owns and controls, utilizing Tasmota on ESP boards, an MQTT broker, Node-RED on a Raspberry Pi, and Grafana for data visualization. By prioritizing local protocols and ensuring that physical switches still function traditionally, Kittleson demonstrates how to build a resilient, agile, and privacy-conscious smart home that provides convenience without sacrificing fundamental reliability.

13h, 19p, 1 comment

Making the Iron Giant Live

The latest edition of the Animation Obsessive newsletter explores the technical artistry behind The Iron Giant (1999), highlighting how director Brad Bird and his team achieved a seamless blend of 3D models and 2D hand-drawn animation. Through innovations like the "Gartoons" system and specific design choices by artists such as Mark Whiting and Steve Markowski, the production ensured the giant's digital presence felt as emotive and "warm" as the traditional animation. Additionally, the newsletter features a profile on Steve Grossman, a self-taught animator who created the music video Look at Me! using Blender, alongside various global animation newsbits covering everything from record-breaking Mexican animated features to upcoming screenings in Britain.

1d, 146p, 64 comments

Show HN: Black Ops II statically recompiled to native ARM64 macOS

The bo2-mac project is an unofficial fan-led initiative that enables Call of Duty: Black Ops II to run natively on Apple Silicon macOS by statically recompiling the game's 32-bit x86 Windows executables into ARM64 machine code. Unlike traditional translation layers like Wine or CrossOver, this method uses Remill to lift code to LLVM IR and implements a hand-written Win32 runtime with Direct3D 11 drawn via Metal. The project does not distribute game files or leaked code; instead, it provides a converter that turns a user's existing Steam copy into a native Mac application. Early testing on an Apple M1 shows the game achieving 51–60 frames per second across the campaign, Zombies, and multiplayer modes against bots, offering a significant performance improvement over virtual machine solutions.

23h, 46p, 6 comments

EventSourcingDB 1.3.0 Is Available

EventSourcingDB 1.3.0 has been released, introducing significant performance improvements, enhanced stability, and a new v2 API. Developed by Golo Roden and the native web, this update features faster EventQL queries and improved write speeds that do not degrade as the database grows. The new API provides more descriptive error codes and handles request outcomes within the response body, while the v1 API will remain supported until January 31, 2027. Key technical enhancements include improved incremental backups, rate limiting by default, and better production-ready features such as OpenMetrics support for the health endpoint. Users are advised to back up their data and update their client SDKs before upgrading, as the new version implements changes to event hashing and data directory ownership.

1d, 266p, 117 comments

AMD's Zen 6 Family: Feeding the Beast

In a detailed technical analysis by George Cozma for Chips and Cheese, AMD’s Zen 6 CPU family is revealed to be a significant evolution of the Zen 5 microarchitecture, characterized by a strategic divergence between the high-frequency Zen 6 and the latency-tolerant Zen 6c cores. Key architectural improvements include a faster Branch Predictor Unit with reduced misprediction latency, an enhanced instruction fetch system capable of allowing a single thread to utilize both decode clusters, and doubled VNNI throughput for improved AI and video processing. Notably, the Zen 6c "dense" cores feature larger out-of-order structures—such as a 640-entry reorder buffer and a larger vector scheduler—alongside a massive 128 MB L3 cache to better absorb memory latency. Performance estimates using SPEC CPU 2026 suggest substantial generational gains, with the EPYC 9686F outperforming previous-generation Zen 5 counterparts by up to 21% in integer tasks, marking a shift in AMD's strategy toward specialized core designs optimized for specific workload behaviors.

2d, 209p, 117 comments

Queens Public Library under ransomware attack

The Queens Public Library is currently managing a significant network outage caused by a ransomware attack, which has disrupted the library's website, mobile app, digital media services, and public computer access at most locations. While all branches remain open as scheduled, certain programs may be impacted, and renewals are temporarily unavailable. Essential services such as free Wi-Fi and book check-out/return machines have been restored across all locations, and public computer access is back online at Central Library, Far Rockaway, Flushing, Glen Oaks, Hunters Point, Jackson Heights, and Kew Gardens Hills. The library has confirmed that no late fines will be charged on borrowed materials during this period.