Hacker News

The Front Page

Vol. I June 30, 2026 Price: A cup of coffee

13h, 782p, 564 comments

Qwen 3.6 27B is the sweet spot for local development

Qwen 3.6 27B is praised as the optimal local AI model for developers, offering strong general‑intelligence capabilities while remaining runnable on consumer hardware. The author details successful tests—creative writing, code generation (e.g., a hexagonal minesweeper), and multi‑token prediction—using llama.cpp with 8‑bit quantization from Hugging Face, achieving 30‑105 tokens/s on a MacBook Max M5 and 50 tokens/s on an RTX 5090. Benchmarks show it comparable to or slightly outperforming frontier models such as DeepSeek‑V4 Flash and approaching GPT‑5‑level performance. The post argues that locally run models enhance privacy, reduce reliance on costly APIs, and foreshadows a future where even larger open‑weight models like GLM 5.2 become affordable for enterprises.

5h, 49p, 16 comments

Exploring PDP-1 Lisp (1960)

PDP‑1 Lisp is a minimalist, interactive Lisp system originally written in 1960 by teenager Peter Deutsch for the DEC PDP‑1, introducing the first read‑eval‑print loop. The page provides a quick‑start guide for bringing up Lisp on a (Pi)DP‑1 replica: setting the Extend switch, loading the lisp.rim tape, configuring memory and stack switches, enabling typewriter input, and basic usage tips (e.g., checking with “nil”, entering expressions like (plus 1 2)). It explains how to load and save Lisp code via paper tape, create custom tapes with encode_fiodec, and use the pdef function to punch out definitions. References to the full PDP‑1 Lisp manual, the original Lisp 1.5 manual, and additional recommended books are given, along with a note on preserving core memory across power cycles and a brief mention of an AI‑tutor approach using markdown files.

12h, 186p, 37 comments

Ornith-1.0: self-improving open-source models for agentic coding

Ornith-1.0 is an open‑source, self‑improving coding agent released by deepreinforce‑ai, available in dense (9B) and Mixture‑of‑Experts (35B, 397B) variants built on top of Gemma 4 and Qwen 3.5. It achieves state‑of‑the‑art results on coding benchmarks such as Terminal‑Bench 2.1, SWE‑Bench, NL2Repo and OpenClaw, outperforming comparable models. The framework uses reinforcement learning to jointly optimize solution rollouts and the scaffolding that guides them, enabling better search trajectories. Licensed under MIT, the models support up to a 256K context window and can be served via vLLM, SGLang or local runtimes (llama.cpp, Ollama) with OpenAI‑compatible APIs, tool‑calling, and agentic coding capabilities. Detailed benchmark tables, deployment instructions, and usage examples are provided in the repository.

19h, 210p, 115 comments

One million passports leaked online

Nearly a million passports and photo IDs from users of Irish cannabis clubs were exposed on public URLs after the clubs’ software provider, Cannabis Club Systems (CCS) (formerly Nefos Solutions), stored the documents without password protection. Security researcher Sammy Azdoufal discovered the flaw, showing that anyone could retrieve ID images, personal details and even private chat messages by altering simple URL parameters. After reporting the issue, CCS temporarily shut down its vulnerable PuffPal app and APIs, began securing the data, and notified authorities, but the response was slow and the company admitted further security lapses. The breach affected users worldwide, including celebrities and thousands from the United States, prompting calls for stricter data‑security practices.

15h, 199p, 32 comments

WATaBoy: JIT-Ing Game Boy Instructions to WASM Beats a Native Interpreter

WATaBoy is a Game Boy emulator that uses just‑in‑time compilation to WebAssembly instead of a native interpreter. By generating Wasm bytecode for each non‑branching instruction block and linking it at runtime via JavaScript, the JIT‑to‑Wasm version runs about 1.2× faster than the native interpreter and ~1.5× faster than the Wasm interpreter, with Safari delivering the best performance across browsers. The project demonstrates Rust‑based Wasm code generation, late‑linking, and indirect‑call dispatch, and shows that even a CPU‑bound emulator can benefit from JIT‑to‑Wasm, suggesting potential for more complex console emulators on platforms like iOS where native JIT is prohibited. Future work includes adding audio, Game Boy Color support, and further PPU optimisations to reduce interpreter fallbacks. © 2026, Seth Humphries.

17h, 235p, 29 comments

What happens when you run a CUDA kernel?

What happens when you run a CUDA kernel – The article walks through the full life‑cycle of a simple vector‑add CUDA program on an RTX 4090, starting with compilation via nvcc (producing PTX, SASS, and a fat binary), then explaining how the host stub registers the kernel and packs arguments into a buffer. It details the lazy upload of the SASS cubin to the GPU, the construction of a Queue Meta Data (QMD) in a pushbuffer, and the doorbell‑ring that triggers the GPU’s host engine to fetch the QMD. Inside the GPU, the compute work distributor schedules warps across 128 SMs, using compiler‑encoded stall counts, yield hints, and scoreboard barriers to manage eligibility and hide memory latency. Memory accesses are coalesced and travel through L1, L2 and GDDR6X, achieving ~80 % of peak DRAM bandwidth. After execution, a completion semaphore signals the CPU, which copies the result back via cudaMemcpy and prints it. Throughout, the post reveals low‑level mechanisms such as ioctl calls, pushbuffer command formats, and SASS control words, illustrating the intricate hardware‑software choreography that underpins a single CUDA kernel launch.

12h, 66p, 19 comments

Micro-Agent: Beat Frontier Models with Collaboration Inside Model API

vLLM Semantic Router introduces a new serving‑layer architecture that turns a single model API call into a bounded collaboration of multiple “micro‑agents.” By routing requests through patterns such as Confidence (escalate only low‑confidence cases), Ratings (controlled parallel ensembles), ReMoM (breadth‑first reasoning with quorum synthesis), Fusion (use disagreement as evidence), and Workflows (budgeted role‑based agent pipelines), the router can dynamically select the optimal loop based on task difficulty, risk, latency, and cost. This “auto‑recipe” approach keeps the external API surface unchanged (e.g., vllm-sr/auto) while improving performance, safety, and cost efficiency, as shown by benchmark gains on LiveCodeBench, GPQA‑Diamond, and Humanity’s Last Exam. The article argues that future AI progress will hinge as much on smarter routers as on frontier models themselves.

53m, 28p, 11 comments

US Supreme Court Just Blew Up EU-US Data Transfers

US Supreme Court decision in Trump v. Slaughter declares the US Federal Trade Commission’s independence unconstitutional, undermining the core “independent” oversight required by EU treaty law for the EU‑US Data Privacy Framework. Since the EU has relied on the FTC’s independence 259 times to deem US data protection “adequate,” the ruling collapses the legal basis of the framework, prompting privacy activist Max Schrems and the group noyb to urge an orderly EU withdrawal and anticipate lawsuits to force the European Commission to repeal the deal. While the formal decision remains in force pending repeal or a CJEU annulment, the ruling casts doubt on the legality of most EU‑US personal data transfers, pressuring companies to reassess reliance on US cloud services, SCCs, and BCRs.

14h, 83p, 64 comments

The Return of Aspect Oriented Programming

Aspect Oriented Programming (AOP) was introduced in the 1990s to let developers treat cross‑cutting concerns—such as logging, security, or error handling—separately from core business logic. The original “join point” model relied on runtime pattern matching, which proved brittle and hard to maintain. The article argues that modern large language models (LLMs) can revive AOP by using separate, reusable documents for each concern (correctness, efficiency, debugability, etc.) and letting the LLM act as a static “weaver” that generates readable code. This approach avoids runtime overhead, leverages the LLM’s semantic understanding to match concerns more robustly, and fits naturally into current AI‑driven code generation workflows.

21h, 210p, 49 comments

Halvar's Guide to Entrepreneurship

Halvar’s Guide to Entrepreneurship is a comprehensive, candid handbook by software‑SaaS founder Thomas Dullien that distills lessons from his two ventures—zynamics (bootstrapped, sold to Google) and optimyze (venture‑backed, sold to Elastic). It covers why people become founders, how to choose and size a target market, and the trade‑offs between bootstrapping and various funding routes. The guide explains venture‑capital dynamics, risk‑return math, and how to run a disciplined fundraise while maintaining momentum and managing investor expectations. It then dives into product strategy (problem‑first ideation, user vs. buyer personas, rapid prototyping), design, hiring, team management, and co‑founder equity splits. Further sections address leadership, culture, employee equity, marketing, sales funnels, and the importance of coaching and personal energy management. Throughout, Dullien stresses honest self‑assessment, clear communication, and the need to align personal motivations with the business’s long‑term trajectory.

20h, 876p, 125 comments

Pollen tried to remove my article and Google is assisting with it

Gergely Orosz reports that after publishing a detailed account of Pollen’s 2022 collapse—highlighting founder Callum Negus‑Fancey’s alleged fraud, unpaid wages, missing pension contributions, a $3.2 million double charge by CTO Bradley Wright, and ensuing lawsuits—Google removed his article from search results following a bogus DMCA claim filed under a fake name from the uninhabited Bouvet Island. Orosz argues the takedown was a coordinated effort by Pollen or its affiliates to erase the story, notes ongoing litigation (e.g., Tayler Ulmer vs. Pollen) seeking back pay and benefits for former employees, and criticizes Google’s handling of fraudulent copyright notices. He vows to continue publishing the expose despite the attempted censorship.

1d, 1072p, 499 comments

GLM 5.2 beats Claude in our benchmarks

Semgrep’s security research team evaluated open‑weight and closed‑source LLMs on an Insecure Direct Object Reference (IDOR) benchmark, finding that their own multimodal pipeline (which adds endpoint discovery and other scaffolding) achieved the highest F1 scores (61% with GPT‑5.5 and 53% with Claude Opus 4.8). Notably, the open‑weight model **GLM 5.2** from Zhipu AI, running with only a basic prompt and no harness, scored 39% F1—surpassing Claude Code’s 32%—while costing roughly $0.17 per vulnerability found. Other open‑weight models lagged far behind. The results highlight that while harnesses significantly boost detection performance, a single open‑weight model can now rival frontier proprietary agents on complex security tasks, offering strong cost and privacy advantages for teams that can run models locally.

1d, 156p, 78 comments

We found a bug in the hyper HTTP library

How we found a bug in the hyper HTTP library – Cloudflare’s Images binding, built in Rust on Workers, began intermittently truncating large image responses after a December 2025 re‑architecture that replaced the FL intermediary with a local Unix‑socket binding. The issue surfaced only under production concurrency when the socket’s outbound buffer filled, causing hyper to issue a shutdown before all buffered data was flushed, resulting in partially delivered images and “end of file” errors in downstream pipelines. Extensive debugging—including reproduction workers, strace traces and kernel‑level analysis—revealed a race condition in hyper’s dispatch loop where the result of poll_flush was discarded, allowing premature connection closure. The fix adds a proper pending check or ensures a flush before shutdown, and has been upstreamed to hyper, stabilizing the Images service and preventing data loss.

11h, 40p, 27 comments

.garden TLD's change to a bad neighborhood

.garden TLD registrations have surged from about 2,500 in 2025 to 147,000 in 2026, with the average risk score climbing from 55 to 84, indicating a sharp rise in abuse. The bulk of the risk is tied to domains using alidns.com nameservers and the Dominet registrar, which push average scores to 87‑94, while other providers like Cloudflare, Spaceship.net, and dnsowl.com show lower risk. Analysts recommend fully blocking the .garden TLD in network defenses, with possible finer‑grained blocks based on registrar or nameserver, as the domain is now considered a high‑risk, abused gTLD.

16h, 24p, 0 comments

Counterexamples in Type Systems

Counterexamples in Type Systems is a curated collection by Stephen Dolan, acknowledging contributions from Andrej Bauer, Leo White, and Jeremy Yallop, that details 31 distinct pitfalls and edge cases in type system design. The list covers issues such as polymorphic references, covariant containers, incomplete variance checking, problems with objects under construction, Curry’s paradox, mutable matching, runtime type misinformation, overloading, subtyping versus inheritance, privacy violations, unstable type expressions, recursion anomalies, scope escape, and various forms of improper subtyping or quantification. Each entry serves as a concrete example to illustrate why certain type system features can lead to unsoundness or unexpected behavior.

18h, 29p, 5 comments

WASM on the JVM Ships Under the Bytecode Alliance

Endive 1.0 has been released under the Bytecode Alliance, providing a pure‑Java WebAssembly runtime for the JVM and replacing the earlier Chicory project. Available on Maven Central, it introduces full support for the WasmGC proposal—allowing garbage‑collected objects to be tracked directly by the JVM—and adds tail‑call optimizations, a tree‑sitter library packaged as a JAR, and host‑integration capabilities that let Java applications invoke Wasm modules seamlessly. The release includes migration guides, documentation, and examples such as compiling javac to Wasm and running it in Endive, with future plans for Component Model support and Cranelift‑based native compilation.

1d, 92p, 19 comments

Mel Brooks is 100 today

Long Live Mel Brooks celebrates the comedian’s 100th birthday, tracing his rise from a Brooklyn‑born, Depression‑era kid who turned comedy into a weapon of self‑preservation to the creator of iconic films such as The Producers, Blazing Saddles, Young Frankenstein and Spaceballs. The piece highlights Brooks’s early influences—from Borscht Belt stand‑ups and a mentorship by Mel Tolkin that introduced him to Gogol’s Dead Souls—and his wartime service, which fueled his disdain for death and shaped his relentless, absurdist humor. It details his partnership with Anne Bancroft, his legendary collaborations with Carl Reiner and Gene Wilder, and his belief that comedy is a protest against mortality, underscoring his lasting impact on American comedy and culture.

16h, 390p, 256 comments

Rocketlab acquires Iridium

Rocket Lab announced a definitive agreement to acquire Iridium Communications for $54 per share in a cash‑and‑stock deal valuing the company at roughly $8 billion. The merger will combine Rocket Lab’s launch and satellite‑manufacturing capabilities with Iridium’s global L‑band spectrum and LEO communications network, creating a fully vertically integrated space firm that can design, build, launch and operate its own constellations. The combined entity aims to expand into satellite IoT, direct‑to‑device services, positioning, navigation and timing, and to secure recurring revenue from Iridium’s 2.55 million subscribers and $871 million 2025 revenue. Completion is targeted for mid‑2027 pending shareholder and regulatory approvals.

5h, 42p, 12 comments

How to corrupt an SQLite database file

How To Corrupt An SQLite Database File outlines numerous scenarios that can lead to SQLite database corruption despite its built‑in resilience. It details risks such as rogue processes overwriting the file, misuse of file descriptors, backups taken during active transactions, deletion or mispairing of hot journal files, and various file‑locking problems (including broken filesystem locks, POSIX lock cancellations, multiple SQLite copies in one app, differing locking protocols, and renaming or linking database files). The document also covers failures to sync writes (e.g., unreliable disk or USB flash sync), hardware faults, memory corruption, OS‑specific issues (LinuxThreads, QNX mmap bugs, filesystem corruption), configuration errors (disabling sync or journal modes), and historic SQLite bugs that could cause corruption. It emphasizes safe practices—using proper backup APIs, avoiding unsafe file operations, ensuring reliable syncs, and keeping SQLite configurations at default protective settings.

14h, 279p, 143 comments

A native graphical shell for SSH

Marcus Lewis proposes a native graphical shell for SSH that serves a home‑screen of apps, each running as a tiny HTTP server accessed via Unix domain sockets and secured by SSH rather than TLS. The “Outer Shell” open‑source project enables these apps—whether simple HTML interfaces or native “outerframe” programs—to discover each other through an API, allowing actions like opening a file in a registered text editor. Lewis demoed the system, linked to documentation for the browser, API, and native app architecture, and argues that this approach could redefine remote server interaction, especially with modern AI‑assisted development making platform‑specific native apps feasible.

2h, 8p, 0 comments

A Fake Shell for Pangenomics

Adrian Sampson presents Flash, a “fake shell” that interprets a small subset of POSIX shell syntax but replaces known pangenomics commands with native Rust calls from his zero‑copy toolkit FlatGFA. By translating scripts into an intermediate representation, Flash can keep data in memory as efficient binary stores, avoid file and pipe I/O, and apply optimizations such as in‑memory BED handling, path‑length shortcuts, duplicate instruction elimination, and automatic use of pre‑converted .flatgfa files. Benchmarks show the same workflow runs up to 28× faster than traditional shell pipelines using odgi, with further gains from optimizations. Sampson argues this approach retains the familiarity of shell scripting while delivering performance comparable to a dedicated API, making FlatGFA more accessible to biologists.

18h, 191p, 31 comments

Dark Sky Lighting

Saving Our Stars advocates dark‑sky lighting—amber, fully shielded fixtures that direct light downward—to curb light pollution and protect safety, health, wildlife, energy use, and the night view of stars. It argues that white LEDs cause glare, disrupt circadian rhythms, harm ecosystems, waste up to 30% of outdoor lighting, and obscure the Milky Way, while amber LEDs are affordable, efficient, and already successful in places like Flagstaff, AZ. The site offers resources, sample letters, and calls for community action to adopt dark‑sky ordinances before further white‑LED installations worsen the problem.

15h, 126p, 41 comments

Working With AI: A concrete example

Carson Gross recounts a recent bug‑fix in the hyperscript parser, using Claude‑AI to pinpoint the regression introduced in version 0.9.91 and to generate tests, but finding the AI’s suggested code fixes either too hacky or overly complex. By leveraging the parser’s existing “follow” mechanism, Gross narrowed the fix to the fetch command, preserving correct behavior for go commands, and manually applied the final patch. He concludes that AI excels at investigation and test creation, yet human expertise remains crucial for crafting clean, maintainable solutions, especially to avoid technical debt and the “Sorcerer’s Apprentice” trap. The piece also reflects on how AI aids older developers by offsetting memory and stamina limits, while warning of potential over‑reliance.

19h, 169p, 44 comments

Sandia National Labs SA3000 8085 CPU

Sandia National Labs SA3000 8085 CPU – In the early 1980s Sandia Lab developed a radiation‑hardened CMOS version of Intel’s 8085, the SA3000, expanding the original 6,500‑transistor design to about 18,000 transistors and fabricating it on 4‑inch wafers with a 3 µm process. Rated for up to 3 × 10⁶ rads (with only a 40% performance loss) it powered critical systems such as the W88 nuclear warhead’s altitude and fuze computer on Trident II missiles, Ball Aerospace’s deep‑space star tracker, and the 1990 CRRES satellite. Support chips (SA3001, SA3002, etc.) were also hardened, and the design was later commercialized by Harris as space‑grade HS1‑80C85RH and military‑grade HS9‑80C85RH parts.

11h, 65p, 30 comments

JumpServer: Open-Source Privileged Access Management

JumpServer is an open‑source Privileged Access Management (PAM) platform that lets DevOps and IT teams securely access SSH, RDP, Kubernetes, databases and RemoteApp endpoints via a web browser. It offers a web UI, web terminal, protocol connectors for graphics, databases, VNC, and facial recognition, as well as remote‑application connectors for Windows and Linux. The project is licensed under GPL‑3.0, has over 30 k stars and 5.7 k forks on GitHub, and provides quick‑start installation scripts for Linux servers. FIT2CLOUD maintains the code, and the repository includes documentation, contribution guidelines, and security policies.

14h, 366p, 108 comments

European ISPs Want Rightsholders Held Accountable for Overblocking Damage

EuroISPA, representing over 3,300 European ISPs, has urged the EU Commission to make rightsholders financially liable for the collateral damage caused by overbroad site‑blocking orders. Citing numerous incidents in Italy, Spain, Belgium and France—where blocking measures have disrupted legitimate services, mail, banking apps and DNS resolvers—the association argues that existing law is already sufficient and that the focus should be on enforcing it rather than expanding blocking obligations. EuroISPA asks the Commission to apply the IPRED framework to require compensation for overblocking and to ease rapid‑block mandates that strain smaller providers.

18h, 385p, 177 comments

Samsung, SK Hynix, Micron Sued in US over Memory Price Fixing

Samsung Electronics, SK hynix and Micron face a U.S. lawsuit filed by 14 consumers and three small businesses alleging that the three DRAM manufacturers colluded to restrict supply and inflate prices by about 700% since 2022, allegedly using a shift to high‑bandwidth memory (HBM) as a pretext. The plaintiffs seek class‑action status, which could expand liability to all buyers of DRAM‑containing products and potentially triple damages if successful. The case follows earlier U.S. antitrust findings against Samsung and SK hynix, though analysts say it is unlikely to affect memory pricing in the near term.

15h, 55p, 30 comments

Venetian Bridge Brawls in 17th and 18th Century Art

Venetian “bridge brawls” were a centuries‑long tradition of factional fist‑fights that turned the city’s many bridges into staged combat arenas. Originating from rival groups—the ship‑building Castellani (“red shrimps”) and the fishermen Nicolotti (“shadows”)—these “battagliole sui ponti” ranged from informal boxing matches to pre‑arranged wars that attracted thousands of spectators. Fights were confined to bridge arches to control violence, with knives and stones banned, while spectators watched from windows, gondolas, and the water below. By the 17th century the spectacle was popular enough to be recorded in paintings and prints, such as works by Joseph Heintz the Younger and Domenico Rossetti. The last major brawl occurred in 1705, after which the tradition faded, leaving its legacy preserved primarily through the art that captured the chaotic, community‑driven contests.

1d, 155p, 22 comments

Dissecting Apple's Sparse Image Format (ASIF)

Erik Schamper details his reverse‑engineering of Apple’s new ASIF (Apple Sparse Image Format), introduced with macOS 26 Tahoe for virtual‑machine disk images. He creates a test file, examines its header (“shdw” magic, version, size, flags, directory offsets, GUID, sector counts, chunk and block sizes), and uses diskimagescontroller binaries to map out the format. The header defines two directory offsets, a 1 MiB chunk size, 512‑byte blocks, and a maximum virtual size just under 4 PiB. ASIF stores data in chunk tables followed by per‑group bitmaps that mark allocated chunks; entries use 55 bits for the chunk number and 9 bits for flags (uninitialized, initialized, unmapped, bitmap). Schamper explains the calculation for locating a chunk from a virtual offset, clarifies the role of reserved entries and bitmaps, and notes that the format’s metadata block contains a plist. He provides a Python implementation in the Dissect project and demonstrates usage with target‑info on a sample ASIF image. The post serves as both a tutorial and a reference for developers working with Apple’s sparse disk images.

16h, 113p, 32 comments

Building Principia for Windows XP

Building Principia for Windows XP details the author’s effort to recreate a fully open‑source build of the 2014 game Principia that runs on Windows XP. By constructing a custom i686‑mingw‑w64 toolchain targeting NT 5.1, patching GCC’s libstdc++ to avoid Vista‑only APIs, and compiling required libraries (curl 8.17.0, FreeType, libjpeg‑turbo, libpng, SDL, zlib) with the new toolchain, the author resolved compatibility issues such as a missing __stdcall calling convention and OpenGL driver constraints. After extensive testing in Wine and on real XP hardware—using the Universal NT Installer for a smooth XP installation—the game boots successfully, albeit with minor font rendering glitches due to driver limits. The project’s scripts and a special XP‑compatible release (2026‑06‑10‑xp) are available on GitHub, though ongoing maintenance will be informal.

9h, 14p, 3 comments

Magicbookshelf.org – a spoiler-aware companion for public domain classics

Magic Bookshelf provides free, public‑domain ebooks accompanied by a “spoiler‑aware” companion tool that lets readers hide or reveal plot details at will. The site features classics such as Crime and Punishment by Fyodor Dostoevsky, Pride and Prejudice by Jane Austen, and The Brothers Karamazov, offering searchable access and chapter breakdowns, along with options for users to suggest additional titles and support the project.

1d, 52p, 80 comments

1.38 Millimeter Microcontroller

MSPM0C1104 is a 24 MHz Arm® Cortex‑M0+ microcontroller with 16 KB flash and 1 KB SRAM, offering a 12‑bit SAR ADC (up to 10 external channels, 1.5 Msps), 5‑V‑tolerant I/Os, DMA, LIN‑UART, SPI, I²C, three timers (14 PWM channels), watchdog, beeper, internal oscillators, CRC‑16, and low‑power modes (run 87 µA/MHz, standby 5 µA, shutdown 200 nA). It operates from –40 °C to 125 °C and 1.62‑3.6 V, with multiple 8‑ to 20‑pin packages. Supported by TI’s LP‑MSPM0C1104 LaunchPad evaluation kit, MSPM0 SDK, and extensive software tools (CCStudio, UniFlash, SysConfig) for rapid development in portable, battery‑powered applications.

4h, 9p, 1 comment

'Humanity has chosen to become idiots': Brown professor discovers mass cheating

Brown University economics professor Roberto Serrano switched his March 2024 midterm to a take‑home, closed‑book format after a campus shooting, aiming to ease student trauma. Instead, 40 of 86 students earned perfect scores, and the class average jumped to 96, prompting Serrano to discover that many answers matched AI‑generated text from ChatGPT. Confronted with the widespread cheating, he gave students a final in‑person exam, after which 27 high‑scoring students dropped the course and the final average fell to 48. Serrano reported the fraud to Brown’s administration, receiving little response, and now bans take‑home exams, assigning no weight to weekly AI‑able homework. He warns that unchecked AI use threatens academic integrity and critical thinking, echoing broader concerns as elite schools grapple with rising AI cheating.

17h, 28p, 2 comments

How we made WINDOW JOIN parallel and vectorized

QuestDB engineers detail how they transformed the costly, multi‑join SQL pattern for averaging bids and asks around trades into a dedicated WINDOW JOIN operator that runs in parallel and leverages SIMD vectorization. By slicing the left‑hand table into page‑frame chunks, each worker fetches a single time‑range slice of the right‑hand table, builds a small in‑memory index, and performs binary searches plus contiguous‑buffer aggregation using AVX2 kernels. This low‑cardinality fast path copies relevant values into per‑symbol buffers, enabling eight‑double SIMD loops for sum, avg, min, max, and count. Benchmarks on a 50 M‑row trades and 150 M‑row prices dataset show the parallel + SIMD WINDOW JOIN is 5× faster than QuestDB’s single‑threaded version and 25× faster than ClickHouse, while Timescale, DuckDB and ClickHouse using rewritten range joins or window functions either exceeded time or memory limits. The article also outlines supported window shapes, aggregates, and future plans such as HORIZON JOIN.

9h, 360p, 94 comments

Free the Icons

Rogue Amoeba urges Apple to restore varied shapes for macOS app icons, criticizing the “Liquid Glass” redesign in macOS 26 (Tahoe) and the forced uniform squircle imposed on third‑party icons, which they say reduces usability and creativity, especially for users with color‑vision deficiencies. They note that macOS 27 (Golden Gate) shows improvements to Apple’s own icons and argue that Apple can easily reverse the restriction, citing internal feedback (FB23388490) and the poor adoption of “Clear” and “Tinted” icon styles. The post calls for “freeing” the icons to allow distinct shapes again.

4h, 40p, 9 comments

Old Computer Challenge

The Old Computer Challenge is a small community of retro‑computing and digital‑minimalism enthusiasts founded by Solène Rapenne in July 2021. Each July a week‑long “OCC” invites participants to run a single‑CPU, low‑memory machine (e.g., 512 MiB RAM, 1 core) and experiment with constraints such as limited internet time or hand‑made creations. The event is coordinated via a mailing list (~tekk/oldcomputer‑challenge@lists.sr.ht), IRC #oldcomputerchallenge on libera.chat, and a website that Tekk now maintains. Past themes include the original 512 MiB RAM challenge (2021), limited internet (2022), “The machine is not timeless” (2023), DIY (2024), and a call for handmade projects in 2026. Participants document their builds and experiences on personal sites, gopher, Gemini, or web pages, and links to many 2025 participants are listed. The community values sharing journals, creative outputs, and preserving the experience of using vintage hardware.

14h, 504p, 234 comments

US Supreme Court rules geofence warrants require constitutional protections

The U.S. Supreme Court ruled 6‑3 in Chatrie v. United States that “geofence” warrants, which compel tech companies to provide smartphone location data for all devices within a virtual perimeter, constitute a Fourth Amendment search and therefore require constitutional privacy protections. Writing for the majority, Justice Elena Kagan held that individuals have a reasonable expectation of privacy in their cell‑phone location records, even when they are in public, and rejected the government’s claim that such data are voluntarily disclosed. The decision affirms that law‑enforcement must obtain a warrant that meets particularity and probable‑cause standards before accessing this information, marking the first major post‑2018 privacy ruling by the Court and a victory for privacy advocates.

14h, 148p, 21 comments

Apple Neural Engine: Architecture, Programming, and Performance

Apple Neural Engine: Architecture, Programming, and Performance by Spencer H. Bryngelson presents a comprehensive reverse‑engineered analysis of Apple’s fixed‑function matrix accelerator, deployed since the A11 iPhone/iPad chips and the M1‑class Macs. The 302‑page study details the datapath, roofline performance limits, energy efficiency, dispatch mechanisms beneath Core ML, compiler and on‑disk program formats, weight‑compression methods, and the kernel driver, firmware, and command protocol. Coverage spans A11–A18 and M1–M5 families, with per‑chip tables and an operation‑by‑device matrix, based on direct measurements from M1 and M5 silicon. Claims are clearly labeled as measured, decompiled, or predicted, and the paper documents the undocumented, version‑fragile user‑space route to the ANE intended for research rather than production use.

5h, 16p, 1 comment

Philae's extraordinary comet landing relived (2024)

Philae made history on 12 November 2014 when ESA’s Rosetta landed the first probe on a comet, 67P/Churyumov‑Gerasimenko, after a ten‑year, 500‑million‑km journey. Chosen to touch down on the smooth Agilkia site, Philae’s descent system failed, causing it to bounce four times before finally settling at Abydos, where it spent 64 hours conducting experiments. Despite the mishap, the lander gathered unprecedented data: seismic readings from the MUPUS hammer, temperature cycles from –180 °C to +145 °C, interior composition via CONSERT, and a suite of organic molecules detected by COSAC and Ptolemy, revealing a highly porous (≈75 %) icy‑dust mixture and non‑magnetic surface. About 80 % of its science goals were met, and the mission’s legacy shapes future ESA comet and asteroid probes such as Comet Interceptor, Hera, Ramses, and M‑Argo.

13h, 128p, 20 comments

Wallace the 6 inch f/2.8 telescope, building it, and hiking with it

Lucas Sifoni recounts a recent outdoor astronomy outing with his custom‑built Wallace, a 153 mm f/2.8 ultra‑wide‑field telescope designed for deep‑sky and nebula imaging. He describes hiking to a limestone cliff, setting up the telescope from his backpack, and capturing ambient video and photos despite the 36 °C heat and a heavy gear load. The post details Wallace’s technical specs: a fast primary mirror, a 2‑inch eyepiece adapter, a 4‑element GSO coma corrector (adjusted to ~79 mm back‑focus), a dual‑module dovetail housing printed in ASA, and carbon‑rod supports. Sifoni shares challenges in grinding the mirror, the corrective optics’ performance, and plans for a lighter, diamond‑generated meniscus mirror in the next iteration. The entry combines personal reflections on the hike with insights for amateur telescope makers.

3h, 8p, 1 comment

Walter S. Arnold–Sculptor/Stone Carver

Walter S. Arnold, a renowned sculptor and stone carver, was featured on the PBS program “Eye on the Arts” in May 2022 and recently received the prestigious Grifo d’Oro award from the City of Genoa, an honor previously bestowed on figures such as Renzo Piano and Peter Gabriel. The site offers galleries of his sculpture, public space commissions, architectural elements, studio information, and contact details for his Chicago‑Italy studio, along with links to interviews, a blog, and a bibliography.

13h, 69p, 7 comments

Ornith-1.0: Self-scaffolding LLMs for agentic coding

Ornith-1.0, unveiled by DeepReinforce in June 2026, is a new open‑source family of LLMs designed for “agentic coding” that ranges from a 9 B dense model suitable for edge devices to a 397 B Mixture‑of‑Experts (MoE) model. Built on Gemma 4 and Qwen 3.5, it employs a self‑improving training framework where the model jointly learns task solutions and the scaffolding harnesses that guide them, enabling automatic discovery of optimal search strategies. Benchmark results show the 397 B version achieving 77.5 on Terminal‑Bench 2.1 and 82.4 on SWE‑Bench Verified, matching or surpassing Claude Opus 4.7 and outperforming comparable open‑source models; even the 9 B variant reaches scores comparable to much larger models. The system includes safeguards against reward hacking through immutable environment boundaries, a deterministic monitor, and a frozen LLM judge. Results across multiple coding and agentic benchmarks demonstrate that Ornith‑1.0 delivers state‑of‑the‑art performance across scales while preserving resource‑efficient deployment options.

14h, 85p, 120 comments

Microsoft Needs Windows Lite

Microsoft Needs Windows Lite – The author argues that Microsoft is losing developers (“Builders”) to macOS and Linux, threatening its software ecosystem. To retain creators and gamers, he proposes a stripped‑down Windows Lite, removing telemetry, ads, AI, and .NET, leaving only core Win32, a lightweight shell, and graphics drivers. Monthly security patches would be optional, and the OS would be priced at $49 per perpetual license without subscriptions. By targeting developers and gamers, the author claims Windows Lite could revive the platform, attract institutional buyers, and become the primary Windows version, ultimately boosting Microsoft’s profitability.

16h, 102p, 108 comments

Is sunscreen the new margarine? (2019)

Is sunscreen the new margarine? An Outside Online long‑read argues that prevailing U.S. sun‑protection guidelines may be misguided, citing research that sunlight (not vitamin D alone) triggers nitric oxide, serotonin and other compounds that lower blood pressure, reduce cardiovascular mortality and improve overall health, while skin‑cancer deaths remain very low. Dermatologists like Richard Weller report that moderate, unprotected sun exposure lowers blood pressure and that sun‑avoidance may carry risks comparable to smoking. Studies from Sweden and Australia show higher all‑cause mortality among sun‑avoiders and endorse balanced sun exposure. Critics warn that sunscreen can block beneficial UV‑A rays, contain hormone‑disrupting chemicals (e.g., oxybenzone), and may not suit darker‑skinned populations who need more UV for vitamin D. The article calls for nuanced, individualized sun recommendations rather than blanket SPF mandates.

13h, 64p, 18 comments

Font-Family Recommendations

Chris Morgan advises web developers to avoid assuming any named font will be available, noting that web‑safe fonts don’t exist and remote fonts can be blocked for security or data‑saving reasons. He recommends always including a generic fallback—especially monospace for code or ASCII art—and using serif or sans‑serif only when needed, while discarding lengthy font stacks that list many system fonts. Morgan suggests limiting named fonts to at most one non‑web font, preferring generic families for consistency and performance, and warns against using system‑ui or ui‑* families for content due to poor language support and misuse. He also highlights quirks like the default size reduction for monospace and urges browsers to improve default font handling.

1d, 969p, 408 comments

HackerRank open sourced its ATS. My resume scored 90/100. Oh wait 74. No – 88

Dan Unparsed tests HackerRank’s open‑source ATS, finding that the same resume receives wildly different scores—ranging from 66 to 99—due to nondeterministic LLM grading, especially in the “projects” and “experience” categories. While technical‑skill checks remain consistent, the tool’s vague rubrics and reliance on AI judgments produce random outcomes, making the system unreliable for fair hiring and potentially discarding qualified candidates by “luck.”

23h, 129p, 30 comments

NUMA: Cores, memory, and the distance between them

NUMA Explained: Why Memory Distance Slows Your VMs details how non‑uniform memory access (NUMA) affects virtual machines, especially under Xen virtualization. It describes NUMA’s evolution from UMA, the performance penalty of remote memory accesses (1.5‑3× slower latency, similar bandwidth loss), and the challenges of memory placement and CPU affinity. The article explains that modern CPUs present multiple NUMA nodes per socket (e.g., AMD EPYC chiplets, Intel SNC), making “one socket‑one node” an outdated model. It contrasts interleaving (evenly spreading memory across nodes) with NUMA‑aware placement, noting interleaving’s predictable but sub‑optimal performance. The piece highlights how Linux kernels expose topology via SRAT/SLIT tables, while Xen’s dom0 traditionally hides NUMA information, causing guests to see a flat memory map and suffer hidden remote accesses. Edera’s recent stack makes Xen NUMA‑aware end‑to‑end, allowing automatic, consistent placement of guest memory and vCPUs without manual tuning, aiming for predictable performance over peak gains.

1d, 967p, 597 comments

Age verification is just a precursor to automated attribution of speech

Age verification laws now being adopted across US states, European nations and Australia are presented as child‑protection measures, but the author argues they are a step toward compulsory attribution of speech, linking online comments to real‑world identities. By requiring users to provide government‑issued IDs, these statutes would give law‑enforcement a direct way to tie “what happened” to “who did it,” bypassing the need for extensive OSINT or subpoenas. The piece warns that such systems could be automated, enabling swift legal or investigative actions against dissenting speech, and urges readers to avoid compliance, suggesting anonymous payment methods like Monero for any necessary verification.

16h, 131p, 63 comments

CachyOS June 2026 Release

CachyOS June 2026 Release introduces the new Hyprland Noctalia desktop option, DNS‑over‑QUIC support, and performance upgrades like extended PGO for Python and a GCC branch‑misprediction patch for modern Intel/AMD CPUs. The installer now includes a preview video, drops paru in favor of Shelly, and adds SDDM for MangoWM, while the audio group gains realtime‑privileges and live‑session improves keyboard layout detection. cachyos‑welcome adds DoQ via blocky, a troubleshooting page, new Azerbaijani, Greek, and French documentation, and various translation updates. chwd gains Turkish localisation, removes handheld packages, resolves multi‑GPU driver conflicts, and adds a 32‑bit Vulkan driver for VMs. cachyos‑settings now enforces 15‑second startup and 10‑second shutdown timeouts for user services, preventing long shutdown delays. Numerous installer fixes address keyboard layout ordering, pacman config copying, and cleanup of leftover files. ISO images are available for desktop and handheld editions across multiple mirrors.

1d, 147p, 94 comments

Herdr: Agent multiplexer that lives in your terminal

herdr is a Rust‑based, terminal‑native agent multiplexer that lets developers run, monitor, and manage AI coding agents (such as Claude, Codex, GitHub Copilot CLI, and others) within organized workspaces, tabs, and panes. It provides mouse‑native pane splitting, real‑time agent state indicators (blocked, working, done, idle), and supports detaching and reattaching sessions without a GUI or Electron layer. Installation is available via a curl script, Homebrew, mise, or direct binary downloads, with preview Windows builds. Users control the interface with tmux‑style keybindings (default prefix Ctrl +B) and can operate over SSH, including remote attach. The tool offers a socket API for custom agent integrations, theming, copy‑mode shortcuts, and session persistence, and is dual‑licensed under AGPL‑3.0 or a commercial license.

5h, 12p, 0 comments

Letos: Create, edit, browse SQLite databases. Formerly known as SQLiteStudio

Letos, formerly known as SQLiteStudio, is a free, open‑source tool for creating, editing, and browsing SQLite databases. The latest major release, Letos 4.0.0 (June 23 2026), introduces a new ERD editor, migration to Qt 6, native ARM64 builds, signed Windows packages, a high‑DPI‑ready interface, and extensive data‑browsing and editing improvements, while remaining portable across Windows, Linux, and macOS. The project is GPL‑licensed, offers no paid edition or telemetry, and provides free code signing via SignPath.io.

11h, 16p, 1 comment

CERN bids farewell to the LHC and enters Long Shutdown 3

The Large Hadron Collider has been switched off to begin CERN’s Long Shutdown 3 (LS3), a massive five‑year programme of maintenance, consolidation and upgrades that will prepare the accelerator for the High‑Luminosity LHC (HiLumi LHC) slated to start in 2030. Since its first beams in 2008, the LHC delivered groundbreaking results—including the 2012 Higgs‑boson discovery—and drove advances in accelerator science, superconducting technology and computing. LS3 will see 1.2 km of magnets and components removed, new cryogenic lines installed, and extensive upgrades to the ATLAS and CMS detectors, including all‑silicon trackers, precision timing and high‑rate calorimeters, to handle up to 200 collisions per bunch crossing. While the machine is offline, thousands of researchers will continue analysing existing data, and the worldwide team of engineers and physicists will transform the LHC complex and related facilities, paving the way for a ten‑fold luminosity increase and new physics opportunities.

17h, 16p, 0 comments

Reading the Internals of PostgreSQL: Database Cluster, Databases, and Tables

Burak Sen outlines PostgreSQL’s internal architecture, describing how a database cluster consists of multiple databases managed by a single instance and identified by OIDs. He details the physical layout of the data directory ($PGDATA), including subfolders for base data, global catalogs, WAL, and tablespaces, and explains how each database gets its own OID‑named subdirectory under base. The article shows how tables and indexes map OIDs to relfilenodes and file paths, noting that VACUUM FULL can change relfilenodes while OIDs remain constant. Tablespaces are covered, with symlinked directories and OID‑based resolution logic. He then examines heap table storage at the page level using the pageinspect extension, describing page headers, line pointers, and tuple layout. The post explains PostgreSQL’s TOAST mechanism for large values, illustrating how oversized data is stored in auxiliary TOAST tables and accessed via pointers. Finally, he demonstrates basic tuple write/read operations, page space management, and index usage, wrapping up a practical walk‑through of sections 1.1‑1.4 of “The Internals of PostgreSQL.”

1d, 253p, 106 comments

The curious case of the disappearing Polish S (2015)

Medium’s editor blocked the Polish letter Ś because its code prevented the default action for Ctrl S to stop the browser’s save dialog, not realizing that on Polish keyboards the Right Alt + S combination (used to type Ś) is internally mapped to Ctrl Alt S. This historic keyboard layout, stemming from typewriter constraints, communist-era hardware limits, and Windows’ Alt‑Gr handling, made Ctrl S clash with the diacritic shortcut. The fix was a simple code change to block Ctrl S only when Alt isn’t pressed, restoring the ability to type Ś and highlighting how legacy language‑specific assumptions can cause modern software bugs.

5h, 87p, 22 comments

Memory Safe Context Switching

Fil-C now provides memory‑safe implementations of the traditional C context‑switching APIs, including setjmp/longjmp and the ucontext family (getcontext, setcontext, makecontext, swapcontext). By encapsulating jump buffers in opaque objects and enforcing strict runtime checks, Fil‑C prevents stack corruption, dangling stacks, and violations of its capability model that normally arise from misuse of these functions. The implementation tracks valid jump targets per stack frame, panics on illegal calls, and integrates with the garbage collector to safely manage roots during fiber switches. Support for ucontext APIs is available from release 0.680 onward (source‑only), employing hidden fiber contexts with controlled lifecycles and thread affinity, ensuring safe coroutine and fiber creation without exposing raw stack pointers.

7h, 43p, 16 comments

Zig – SPIR-V Backend Progress

Zig’s development in 2026 has focused on major compiler and tooling upgrades. The SPIR‑V backend was overhauled with a new @SpirvType builtin, calling‑convention‑based execution‑mode handling, multi‑threaded codegen, and object‑file linking, boosting test coverage and making shader development feasible. The LLVM backend received a redesign of @bitCast semantics—now reinterpreting logical bits rather than memory layouts—plus integer lowering improvements that yield ~5 % faster Zig compilation. A new ELF linker adds fast incremental builds and work toward DWARF support. Zig’s build system was split into a lightweight “configurer” and an optimised “maker”, dramatically cutting build times. Incremental compilation was enabled for LLVM, and a massive type‑resolution rewrite improved lazy analysis, dependency‑loop diagnostics, and overall incremental performance. std.Io gained experimental io_uring and Grand Central Dispatch backends, while package management now stores fetched packages locally in zig‑pkg and supports a --fork flag for easy overrides. On Windows, Zig prefers native NT APIs over kernel32 wrappers, reducing overhead and failures. Finally, the “zig libc” project is replacing vendored C sources with Zig wrappers, cutting repository size and enabling cross‑module optimisations.

8h, 63p, 6 comments

Kb – Prolog Knowledge Base

kb‑prolog is a prototype hyper‑relational knowledge base that stores information as statement(Subject, Predicate, Object, Properties) triples, allowing subjects and objects to be nested statements for reification. It combines content‑addressable storage (files are SHA‑256 hashed and committed atomically with graph metadata), automatic deduplication, and time‑travel versioning via linked statement replacements. Built with Trealla Prolog as the primary runtime, it accesses SQLite and Raylib through C FFI libraries, and includes an interactive Raylib‑based GUI for graph visualization, image previews, and query input. The repository provides source code, build scripts (Clang, X11/macOS), submodules for dependencies, and a Nix shell for reproducible builds, along with command‑line tools for loading contexts, asserting statements, searching, managing CAS objects, and performing maintenance tasks. The project is licensed under GPL‑3.0.

7h, 194p, 107 comments

South Korea to spend $1T on more memory chip production and humanoid robots

South Korea announced a $1 trillion investment plan to boost memory‑chip production, build AI data centers, and commercialize humanoid robots by 2028. Samsung and SK Hynix will spend $585 billion on new DRAM fabs to double output, while SK Group, GS Group and Naver will fund $357 billion for nationwide AI data centers. Hyundai Motor, together with its subsidiary Boston Dynamics, is allocating $5.8 billion to mass‑produce up to 30,000 Atlas humanoid robots annually and aims to deploy them across ten major industries, training 10,000 AI‑robotics specialists. The projects raise concerns over energy and water demand, labor‑union opposition to robot deployment, and calls for profit‑sharing from the booming chip sector.

22h, 494p, 296 comments

30-year sentence for transporting zines is a five-alarm fire for free speech

Daniel “Des” Sanchez Estrada received a 30‑year federal prison sentence for transporting a box of anarchist‑topical zines, despite not participating in the July 4, 2025 protest at the Prairieland immigration jail where a police officer was shot. Prosecutors tied the zines to his wife Maricela Rueda’s alleged extremist ideology, using the Trump administration’s NSPM‑7 “counter‑terrorism” memo to treat possession of the pamphlets as criminal evidence. The case, part of a broader crackdown that also targeted journalists like Don Lemon and independent reporter Georgia Fort, signals a widening effort to criminalize dissent and information sharing, raising alarms about First Amendment erosion under the current administration.

5h, 39p, 9 comments

Alan Kay on the meaning of "object-oriented programming" (2003)

Alan Kay explains that his original concept of object‑oriented programming was rooted in messaging between autonomous entities, the local retention and protection of state‑process within each object, and an extreme degree of late‑binding. He coined the term around 1967, drawing on ideas from biology, the Burroughs B5000 architecture, and early computing systems such as Sketchpad and PLANNER. Kay downplays inheritance, polymorphism and encapsulation as later additions, emphasizing instead a “non‑data” approach where objects communicate solely via messages, making the traditional data‑container view of objects insufficient for modeling causality over time.

6h, 25p, 9 comments

Open Memory Protocol – One Memory Store for Claude, ChatGPT, Curso

Open Memory Protocol (OMP) is a vendor‑neutral, open‑source standard that lets AI tools share and persist user memories across applications, sessions, and devices. It defines a precise memory object schema, storage format, and HTTP API, and includes a reference Docker‑based server, TypeScript and Python SDKs, and adapters for Claude, OpenAI, Cursor, and others. Users can run a self‑hosted OMP server, connect tools via simple configuration or the OMP Bridge browser extension, and programmatically save, search, and hand off conversations, ensuring memories remain under user control and avoid siloed AI contexts. The project is licensed under Apache 2.0 and invites community contributions for adapters, SDKs, and spec improvements.

7h, 50p, 4 comments

SQLite improving performance with pre-sort

The article by Anders Murphy explores how SQLite performance can be dramatically enhanced by pre‑sorting large batches of random BLOB primary‑key inserts, showing that applying a custom 8‑byte unsigned comparison—mirroring SQLite's memcmp ordering—raises insert throughput from roughly 100 k rows per second to about 2–2.5× faster, thereby illustrating the value of batch processing and ordering for unordered random data.

7h, 96p, 70 comments

Scientists find molecular-level evidence for two structures in liquid water

Scientists provide molecular-level proof that liquid water comprises two distinct local structures: using unsupervised deep‑learning analysis of 74 million configurations from large‑scale molecular dynamics simulations (TIP4P/Ice model), researchers identified a denser, disordered Structure A and a less‑dense, ordered Structure B. The two structures interconvert via different pathways that change topology near the predicted liquid‑liquid phase boundary, supporting the long‑standing two‑state model and suggesting a liquid‑liquid phase transition in supercooled water. The findings, published in Nature Physics (2026), offer a clear structural signature for the hypothesized high‑density and low‑density liquid phases and open avenues for experimental verification and broader implications for biology and geology.

18h, 106p, 56 comments

Rebuilding the Computer Room

Alex Chan reflects on how the once‑fixed “computer room” of his childhood has vanished as laptops, smartphones and wearables made computing portable, freeing us from desks but also flooding our lives with constant, addictive notifications. He describes a personal experiment to rebuild those boundaries: strict notification limits, using a desktop as his primary computer, keeping his phone on a stand in a dedicated office, wearing non‑pocket clothing at home, and trying a screen‑less fitness tracker. The resulting reduction in digital distraction has made him calmer and more focused, illustrating a growing trend to re‑introduce physical limits on technology to protect attention and well‑being.

13h, 116p, 90 comments

The Radiation Exposure Lie

How to lie about radiation examines the myth that low‑level, chronic radiation exposure is a major health threat, arguing that evidence for such harm is weak and often misinterpreted. It reviews the Chernobyl disaster, showing that despite massive fallout, only a few hundred early deaths occurred and cancer rates in exposed populations remained low. The article then analyzes the Taiwanese cobalt‑60 apartment case, noting that studies claiming increased cancer risks rely on questionable statistical methods and that overall cancer incidence was actually lower than in the general population. It highlights the INWORKS cohort of 300,000 nuclear workers as the best evidence of a modest risk—about a 5 % increase in cancer mortality per 100 mSv—yet stresses that this effect is tiny compared with other risk factors like smoking. The piece concludes that stringent regulations aimed at preventing negligible radiation doses impose unnecessary costs, while high‑dose incidents, not low‑level exposure, are where genuine danger lies.

7h, 20p, 6 comments

macOS Golden Gate icon comparison

The article discusses a visual comparison between macOS Golden Gate and Tahoe, highlighting updated icon designs, bolder colors, and changes in the Liquid Glass effect. It notes particular shifts in the Finder icon and the overall sharpness or refinement of the appearance, suggesting this may be an early-beta change before the final launch. The comparison focuses on how these updates affect user experience and the visual aesthetics of the operating system's interface.

17h, 100p, 85 comments

NixOS 26.05

NixOS 26.05 "Yarara" was released by release managers yayayakak and jopejoe1, offering seven months of bugfixes and security updates until the end of 2026. The previous version, 25.11 "Xantusia", is now deprecated and will no longer receive support after June 30, 2026. The release includes 85 new modules, 20442 updated packages in Nixpkgs, and systemd integration in stage 1 boot images. Key changes also involve dropping support for x86_64-darwin systems in future releases. The update to GNOME 50 "Tokyo" and GCC 15 are highlighted, alongside gratitude to 2842 contributors for the collaborative effort.

15h, 49p, 8 comments

Decker Fantasy Camp 2026

Decker Fantasy Camp 2026 is a month‑long game jam on itch.io running from July 1 to August 1, 2026, inviting creators to craft any type of project using the Decker engine and submit it as a web‑playable HTML export. Participants may produce visual novels, poems, printable zines, arcade games, utilities, or any other multimedia work—chickens are optional but welcomed. Submissions must be original, not generated by AI, and may incorporate public‑domain or Creative Commons assets; teams and multiple entries are allowed. Resources include the Decker Community forum, an interactive tutorial by Phinxel, the GitHub repository for bug reports, and official documentation, with links to past Decker jams for inspiration.

11h, 15p, 0 comments

wolfSSL vs. MbedTLS

wolfSSL outperforms MbedTLS in an apples‑to‑apples benchmark across Intel i9‑11950H, Raspberry Pi 5 (Cortex‑A76), STM32H563 (Cortex‑M33) and Microchip PolarFire (RISC‑V) platforms, using wolfSSL v5.9.1 and MbedTLS 3.6.6 (June 2026). Hand‑written assembly and single‑precision math in wolfCrypt deliver 15‑53× faster public‑key ops and 3‑28× faster symmetric ciphers, with AES‑GCM up to 17× faster on Intel and 4× on ARM. wolfSSL also provides a complete post‑quantum suite (ML‑KEM, ML‑DSA, SLH‑DSA, LMS, XMSS) and many extra algorithms that MbedTLS lacks, along with FIPS 140‑3, DO‑178C, ISO 26262 certifications, PSA Crypto API support and commercial 24/7 backing.

11h, 19p, 7 comments

ACL 1.0: A source-available commercial license for the AI era

Auditable Commercial License (ACL) v1.0 is a source‑available commercial license tailored for the AI era, allowing users to read, audit, and modify code internally while prohibiting redistribution and hosted services. It introduces three key features: auditability (explicit rights to inspect source for security and compliance), an AI‑training protection clause that blocks use of the licensed work as training data except under a safe harbor, and a self‑expiring mechanism that automatically converts the license to Apache 2.0 after four years or accelerates this conversion if a continuity event (e.g., vendor bankruptcy) occurs. The site offers a generator to customize the license with details such as licensor, jurisdiction, and release dates, and provides comparison tables against other licenses like Apache 2.0, BUSL 1.1, and Elastic 2.0. The full license text is released under CC0 1.0, but users are warned to seek legal counsel before deployment.

18h, 51p, 43 comments

Type-checked non-empty strings

Haskell koan: Type‑checked non‑empty strings describes a technique used at Bellroy to enforce at compile time that certain text values are never empty. By leveraging GHC 9.10’s RequiredTypeArguments and custom type‑level constraints, the author defines a NonEmptyText constructor that yields a type error when given an empty literal, replacing thousands of Template Haskell splices and cutting build time by about 10 %. The post explains the implementation details—using a type family IsNonEmptySymbol, an overlapping class instance to emit a custom error, and the required language extensions—and shows how the same pattern can validate other predicates such as positive naturals or DynamoDB table names. It concludes with remarks on the growing power of dependent‑type features in Haskell.

1d, 484p, 176 comments

Librepods: AirPods liberated

LibrePods is an open‑source project that reverse‑engineers Apple’s proprietary AirPods protocol to enable features such as noise‑control mode switching, ear detection, battery status, head gestures, conversational awareness, hearing‑aid settings, and multi‑device connectivity on Linux and Android. The README lists which features work on each platform, explains required steps like VendorID spoofing, and notes future plans (e.g., Find My integration, spatial audio, heart‑rate monitoring) that may need root access. Development is primarily in Kotlin with AI‑assisted code generation, and the project is licensed under GPL‑3.0, with a large contributor community and several related alternatives for Windows and Steam Deck.