Image

Hacker News

ImageThe Front PageImage

Vol. I June 9, 2026 Price: A cup of coffee

7h, 356p, 322 comments

Apple reveals new AI architecture built around Google Gemini models

Image

Apple announced a major overhaul of Apple Intelligence, introducing a new AI architecture built around Apple Foundation Models co-developed with Google using technology behind the Gemini family. The upgraded system is designed to run both on-device and through Private Cloud Compute, with a new orchestrator coordinating features across Apple platforms based on the active app and user task. Apple said the collaboration enables stronger reasoning, multimodal support, realistic image creation, advanced photo editing, visual question answering, and, on some unspecified devices, speech generation, better dictation, and improved natural language understanding. The company also emphasized privacy, saying user data is used only to fulfill immediate requests and is not accessible to Apple or third parties.

10h, 503p, 350 comments

MiMo-v2.5-Pro-UltraSpeed: 1T model with 1000 tokens per second

Xiaomilaunched the MiMo‑V2.5‑Pro‑UltraSpeed API, allowing a 1T‑parameter model to decode at 1000 tokens/s for the first time; the service is offered at three times the regular price but delivers roughly ten‑fold speed using only commodity GPUs via extreme model‑system codesign with TileRT, employing FP4 quantization and the novel DFlash speculative decoding technique to dramatically raise acceptance rates and enable millisecond‑level response for tasks such as real‑time coding agents, anti‑fraud detection and medical imaging analysis; approved users can access a free chat trial from June 9 to June 23 2026, subject to a daily queue limit of 10 entries, a 30‑minute session cap, and automatic session release after 5 minutes of inactivity.

11h, 407p, 323 comments

xAI is looking more like a datacentre REIT than a frontier lab

Image

xAI, now part of SpaceX following their February merger, has shifted focus toward functioning more like a data center real estate investment trust (REIT) with a frontier lab attached, following recent high-profile deals with Anthropic and Google. These partnerships provide xAI access to over $2 billion in monthly revenue through capacity agreements, including 300 MW of computing power (approximately 220k GPUs) for Anthropic and 110k GPUs for Google. While legal pressures—such as Elon Musk’s ongoing disputes with OpenAI—and shareholder incentives (Google being a SpaceX investor) may influence the deals, the core driver appears to be xAI’s operational expertise in rapid datacenter construction. SpaceX/xAI built its first Colossus 1 datacenter in Dallas in 122 days, a pace far faster than traditional hyperscalers, giving it a strategic advantage amid global compute shortages. Analysts note the deals likely involve newer GPUs like the GB200, though specifics remain unconfirmed. Despite skepticism about financial engineering tied to SpaceX’s upcoming IPO, the agreements reflect real infrastructure value, as even older hardware remains critical amid shortages. Meanwhile, Grok’s development faces uncertainty due to reallocated resources, but surplus capacity may be monetized. The success of this venture hinges on balancing these dynamics, positioning xAI as a hybrid infrastructure-play and AI lab in a competitive, capital-intensive market.

8h, 156p, 286 comments

Ask HN: What are tools you have made for yourself since the advent of AI?

A recent Hacker News discussion asks users to share tools they have built for personal use since the rise of AI. One contributor mentions creating tools without relying on AI, instead using code generators and minimal toolchains for fast compilation on low-resource hardware. Another respondent describes a custom harness built with Dagger offering features like diff and time travel, alongside a markdown search system powered by Typesense to enhance personal knowledge management. The latter project is part of a broader effort to achieve more efficiency and better outcomes through tailored software solutions.

10h, 411p, 436 comments

AI is slowing down

Image

The article argues that the AI industry, driven by companies like OpenAI and Anthropic, is trapped in an unsustainable financial bubble requiring unprecedented revenue growth to justify massive infrastructure investments. Ed Zitron contends that AI demands over $3 trillion in data center spending and must generate $2 trillion in annual revenue by 2030 to sustain itself, with Anthropic and OpenAI needing to grow revenues from $60 billion in 2026 to $400 billion by 2030—a 496% increase—while also raising hundreds of billions in funding. He highlights the lack of external demand for AI compute, the risks of token-based billing without cost controls, and the "giant metal spider" metaphor illustrating AI's inefficiency and destructive costs. The piece critiques the industry's focus on growth over profitability, the desperation of companies burning through subsidies, and the potential for a crash as financial realities clash with hyperbolic promises from tech leaders and venture capitalists.

11h, 142p, 20 comments

OCaml Onboarding: Introduction to the Dune build system

Image

The articleOCaml Onboarding: Introduction to the Dune build system explains that Dune serves as the primary build tool for OCaml projects, using a dune-project metadata file and per‑directory dune configuration to declare stanzas for libraries, executables, and tests, and provides commands such as dune build, dune exec, dune runtest and dune init for building, running, testing and scaffolding new OCaml projects like the toy helloer example.

9h, 584p, 235 comments

Stop the Apple Music app from launching

Image

Music Decoy silently stops thesystem Music app from opening whenever the ▶ Play key is pressed, by presenting the same bundle identifier com.apple.Music and running as a lightweight process; it only activates when no app is already playing audio, or when a Bluetooth headset triggers playback or a call ends. Users can configure it to launch another app—such as Spotify—via a defaults write command, and stop it by quitting through Activity Monitor or killall ‘Music Decoy’. Alternatives include unloading the rcd daemon or using tools like noTunes, but they either disable the Play button or consume a small amount of CPU.

13h, 97p, 73 comments

Life is too short for a slow terminal

Image

Mijndert Stuij explains how he keeps his terminal lightning‑fast by stripping away heavy frameworks, using only three lightweight Zsh plugins and lazy‑loading costly tools like nvm and kubectl. He caches completion data, employs an asynchronous prompt (Pure) to avoid git‑status delays, and runs his shell in the GPU‑accelerated Ghostty emulator. Stuij provides concrete commands for measuring startup time, profiling with zprof, and identifying bottlenecks, urging users to keep configurations minimal and intentional so that opening new tabs and typing feels instant. All settings are available in his public dotfiles repository.

8h, 133p, 22 comments

Full Reverse Engineering of the TI-84 Plus Operating System

The TI-84 Plus OS reverse engineering notes detail the system architecture of the ti84plus.rom (2.55MP), a Z80-based single-tasking monitor with 1MiB flash and 128KiB RAM managed via a 4-slot paging scheme and bcall system for cross-page code access. The OS centers on four pillars: paging/bcall mechanisms (enabling 64KiB addressable space), the floating-point engine (9-byte BCD reals in OP registers), the variable system (Variable Allocation Table for named objects), and the tokenizer/parser (TI-BASIC token execution). I/O subsystems include IM1 interrupts, LCD drivers, keypad scanning, and link ports. Documentation spans 14 core subsystem pages (e.g., memory map, paging, interrupts, VAT) and deep-dive topics (calculation engine, graphing, TI-BASIC, I/O), with confidence levels assigned to details based on verification status in disassembly.

4h, 9p, 0 comments

Expanding Private Cloud Compute

Apple is expanding its Private Cloud Compute (PCC) infrastructure beyond its own data centers, partnering with Google and NVIDIA to run Apple Intelligence workloads on Google Cloud while preserving the same stringent privacy and security guarantees originally established for PCC on Apple silicon. The new deployment leverages NVIDIA Confidential Computing with GPUs, Intel TDX on CPUs, and Google’s Titan chip to create a confidential inference pipeline that incorporates end‑to‑end transparency, stateless computation, no privileged runtime access, non‑targetability, and verifiable transparency. Apple retains full control over PCC software, which devices will only trust if cryptographically approved by Apple, and the company will publish all binaries, provide public research tooling, and offer live PCC nodes in research mode through its Security Bounty Program. The rollout will ramp up through a summer preview, with further technical details to be shared at the Confidential Computing Summit and in an updated PCC Security Guide later this year.

1d, 883p, 400 comments

Building from zero after addiction, prison, and a felony

Image

Gavin Ray recounts his turbulent path from a 14‑year‑old drug dealer in Colorado’s maximum‑security juvenile prison to a sober software engineer at Hasura (now PromptQL). After stints in landscaping, college, and a county jail where a tech internship opportunity emerged, Ray’s career was derailed by relapse and repeated job rejections due to his felony record. A turning point came when he discovered Hasura’s GraphQL engine, became an active contributor, and was eventually hired—offering him a paycheck and a chance to work on a product he truly valued. Ray’s narrative underscores the harsh reality of navigating addiction, incarceration, and systemic bias while highlighting the pivotal role of mentorship and second chances in rebuilding a life and career.

6h, 12p, 0 comments

Show HN: HTTP/3 and raw QUIC client/server APIs for Node.js

Image

The currentspace/http3 repository on GitHub provides a Node.js package (@currentspace/http3) for implementing HTTP/3, HTTP/2, and raw QUIC server/client functionality, powered by Rust and the quiche library. The package supports HTTP/3 over QUIC/UDP with HTTP/2 fallback over TLS/TCP, features explicit runtime selection modes (fast, portable, or auto) with platform-native I/O optimizations including kqueue for macOS and io_uring for Linux, and includes adapters for fetch/SSE/EventSource as well as Express compatibility. Version 0.6.0 added mutual TLS support for raw QUIC clients and servers, explicit client certificate policies, and peer certificate inspection capabilities, with prebuilt native binaries available for Linux x64/arm64 and macOS arm64 platforms. The project is MIT licensed and includes comprehensive documentation covering runtime modes, deployment matrices, and benchmarking utilities.

18h, 197p, 184 comments

Age verification tech could put children at greater risk, says think tank

Image

The Foundation for Information Policy Research (FIPR) warns that UK government proposals for mandatory age verification on social media and online services could endanger children by exposing them to risks like blackmail and data misuse, while failing to address harmful content or addictive app design. FIPR argues that technical solutions such as facial recognition or biometric data collection may exclude vulnerable adults, marginalize disadvantaged groups, and push underage users to riskier, unregulated platforms. The think tank criticizes the Children’s Wellbeing and Schools Act 2026 for granting sweeping regulatory powers without adequate privacy protections, and proposes instead a system of content tagging and blocking akin to film ratings, allowing parents to control access. FIPR also highlights that online age verification differs fundamentally from in-person ID checks, as it creates digital data trails, and warns that such measures may not curb the viral spread of harmful content or protect freedom of expression, urging regulators to prioritize tackling root causes rather than relying on unproven technologies.

9h, 17p, 4 comments

Show HN: Startup sci-fi novel that took me 5 years to write

Blockchained by Milo C. Kingston follows Terry, a newly promoted VP of Engineering who quits his job rather than join his firm’s data‑mining scheme, poaches his team, and launches a blockchain‑based social platform promising user privacy and an ad‑free experience; after two years the venture still hasn’t launched and he faces mounting financial pressure when he flies to Hong Kong for a lifeline investor who turns out to be linked to the Chinese government, turning the story into a technothriller that pits ambition against integrity while examining capitalist exploitation, AI, cryptocurrency, censorship, and the East‑West political fault line in Hong Kong—reviews liken it to Eggers, Hart, and Doctorow, calling it an authentic and “5‑star” read that uniquely blends startup drama with geopolitical suspense.

1h, 36p, 6 comments

Hermes Agent – Open-source AI agent with persistent memory

Image

这篇内容介绍了 Hermes Agent 是一个开源人工智能代理，具备持续记忆、自动技能创建和多平台联通能力，适合MLOps与AI训练。它可以在服务器上运行，保留用户数据，不需要外部依赖，并支持跨平台连接如 Telegram、Slack、Discord等。Hermes 提供了丰富的技能系统和快速自动化操作，适合处理复杂任务与长期任务。について简要说，作者推荐使用可轻松安装并长期使用的开放源码项目。

7h, 110p, 52 comments

Why are cells small?

Image

Surface area-to-volume ratio and diffusion are the primary physical constraints explaining why cells are small. As cells grow, their volume increases faster than surface area, limiting the membrane's ability to import nutrients and export waste efficiently. Meanwhile, diffusion rates decrease with size, as molecules must travel farther to interact—collisions that drive cellular processes occur more slowly in larger volumes. Despite evolutionary adaptations like specialized shapes (e.g., biconcave red blood cells) and compartmentalization in eukaryotic cells, cells must balance these constraints. Exceptions like Thiomargarita magnifica bacteria and bubble algae achieve unusual sizes through specialized adaptations such as creating vacuoles that concentrate molecules at the cell periphery, effectively reducing diffusion distances.

5h, 30p, 20 comments

Doing something that’s never been done before (2025)

Tal Globus argues in his piece Doing Something That’s Never Been Done Before that to maximize the odds of being the first to tackle an idea, one should pursue projects that are obscure, time‑consuming, and difficult, especially those featuring progressive dependency where each step only becomes apparent after the previous one is completed, thereby drastically reducing the pool of potential imitators and increasing the likelihood of originality.

9h, 288p, 49 comments

Massachusetts bans sale of precise location data in new privacy rights bill

Image

Massachusetts lawmakers unanimously passed the Consumer Data Privacy Act, banning the sale of precise location data and protecting biometric, religious, immigration, and sexual‑orientation information for companies that handle data of more than 100,000 consumers, with the bill now headed to the governor’s office for expected signing; privacy advocates including Evan Greer of Fight for the Future and the ACLU hailed it as a landmark step against Big Tech surveillance.

16h, 136p, 196 comments

Replies to comments on my "LLMs are eroding my career" post

Image

The author, an engineer piloting LLMs full-time, expresses mixed emotions about AI's impact on their career. While acknowledging LLMs' limitations in handling domain-specific knowledge (e.g., local tax regulations, accounting workflows, ledger details), they note that newer models and agent-friendly tooling now handle tasks once requiring years of human expertise. This shift reduces reliance on senior colleagues, raising anxiety but also creating opportunities to leverage AI for adversarial code reviews, E2E testing, and strategic delays to mitigate risks. The author critiques companies prioritizing speed over thorough documentation and warns of AI's potential to commoditize technical professions, drawing parallels to copywriting and UX writing, where LLMs have drastically cut demand for human roles. They argue that while elite experts may remain relevant, most workers face obsolescence as AI improves. Despite skepticism about industry "FUD," the author urges vigilance against dismissing AI's unique, rapid advancements compared to past tech waves, emphasizing the need to adapt without fatalism.

13h, 119p, 75 comments

Spanish traders set the standard for GnuCash database design

The webpage discusses how historical influences shaped GnuCash’s database design, citing the impact of Spanish traders and the need to avoid fractional counting. It highlights a shift from using minor units to decimal precision, illustrating trade-offs between old methods and modern systems. The content emphasizes that what seems odd at first can become a practical solution, especially for managing commodities and historical currency systems.

19h, 413p, 89 comments

How much of Thermo Fisher's antibody data has been manipulated?

Image

As of June 3, 2026, researchers Reese Richardson and Sholto David have identified over 450 manipulated images in Thermo Fisher Scientific's online antibody catalog, including 50 instances of duplicated background patterns across different products, with similar issues also documented for Abcam (+1). The analysis, detailed in a blog post and a public Zenodo repository, reveals Western blot verification data where bands are identical after rotation or flipping, images display "brushstrokes" suggesting digital alteration, and repetitive background noise indicating copy-pasted regions. These manipulations raise concerns about the reliability of antibody validation data critical for biomedical research, as antibodies failing specificity or selectivity can lead to irreproducibility. Thermo Fisher claims its data is internally or third-party generated, but the findings highlight risks for scientists relying on vendor-provided evidence, with individual antibodies costing $400–$500. The researchers urge systematic validation and invite contributions to their repository to expand scrutiny of vendor practices.

10h, 165p, 175 comments

Why are so many young people getting cancer?

Image

Young adults worldwide are experiencing a notable rise in cancers once predominantly seen in older populations, including colorectal, uterine, and liver cancers, with over 9,000 diagnoses daily in those under 50. While researchers have identified potential factors such as obesity, ultra-processed foods, and environmental exposures contributing to this trend, the exact causes remain unclear. Colorectal cancer rates in the U.S. have climbed steadily since 2010, becoming the leading cause of cancer death in adults aged 20-49. Some increases may stem from changes in cancer detection or classification, like pancreatic cancer definitions in the 2010s, but other cancers show genuine rising incidence tied to "birth cohort effects," where specific generations face higher risks. Obesity is linked to some of these cancers but doesn’t fully explain the trend, prompting calls to investigate novel exposures. Researchers warn that today’s elevated risks in younger populations could translate to higher cancer rates in older age groups as this generation ages, underscoring the urgency to understand and address these shifts.

4h, 21p, 2 comments

All the Ways Europe Is Ditching American Technology

Image

European governments and organizations are accelerating efforts to reduce dependence on U.S. technology firms amid concerns over data control, geopolitical tensions, and the Trump administration’s policies. Since the start of President Trump’s second term, European entities have moved to adopt open-source alternatives and local solutions, including France’s Qwant search engine, LaSuite open-source office software, and Euro-Office, a collaborative European document platform. The European Parliament replaced Google with Qwant, while Dutch, French, and German cities shifted away from Microsoft Office and Google Docs. Other initiatives include migrating code repositories from Microsoft-owned GitHub to national alternatives and Finland’s rejection of AWS for election data. The push, driven by fears of data access under U.S. laws like the CLOUD Act and FISA, as well as closer ties between Big Tech and the Trump administration, reflects a broader drive for “digital sovereignty.” Despite these moves, experts caution that U.S. firms dominate Europe’s digital infrastructure, making a complete break challenging, though European officials stress the urgency of the transition in response to evolving geopolitical dynamics.

22h, 125p, 33 comments

Richard Scolyer Has Died

Image

Australian pathologist and melanoma researcher Dr. Richard Scolyer, 59, has died three years after being diagnosed with an aggressive glioblastoma brain tumor, opting for a world-first experimental immunotherapy treatment developed by his friend and colleague Professor Georgina Long based on their groundbreaking melanoma research. Scolyer, who co-directed the Melanoma Institute Australia and was named Australian of the Year in 2024 for his work improving melanoma survival rates from less than 10% to 50%, became the first brain cancer patient to receive pre-surgery combination immunotherapy and a personalized tumor vaccine, with subsequent scans showing a positive immune response that has prompted a small clinical trial in the US. Known for documenting his treatment journey publicly, Scolyer called for continued scientific bravery and government funding for medical innovation, stating his belief in "changing the future for others," while leaving behind a legacy as a pioneering scientist who "took us into his confidence" during his courageous battle with the disease.

2d, 75p, 105 comments

Splash Is a Colour Format

Image

It seems your message contains some complexity or may reference specific context I'm less entirely aware of. To better assist you, could you clarify your request or provide additional details? For instance, are you looking for help with a particular task, information, or guidance? I'm here to support you when you let me know what you need. Feel free to share more context so I can respond effectively. If you have a specific question or task in mind, please let me know! 😊 — Your Assistant

17h, 76p, 10 comments

A Family Project (2022)

Image

Tess Durrell administers to her deceased mother, Lola Weldine, preparing her for burial at home on family land in rural North Carolina. With her three brothers, she follows legal guidelines for home burials, constructing a casket from oak lumber provided by neighbors and digging a 6-foot-deep grave on her 40-acre property near Hanging Rock. The family inters Lola, an 80-year-old former schoolteacher and passionate gardener, in a red silk-lined coffin amid flowers from her beloved vegetable garden. The intimate, DIY funeral—marked by absence of a officiant—reflects Lola’s love of music, religion, and nature, with the family adhering to her wish to be buried on their ancestral land. One year later, the piece underscores the logistical challenges and emotional catharsis of reclaiming end-of-life agency, positioning the South as a model for personal, non-commercial death rituals. Keywords: home burial, family-led funeral, casket construction, North Carolina death rites, environmental burial

5h, 7p, 1 comment

The Third Generation of Apple's Foundation Models

Image

Apple has unveiled its third generation of Apple Foundation Models (AFM), a family of five AI models developed in collaboration with Google that power the new Apple Intelligence system. The family includes two on-device models—AFM 3 Core, a 3-billion-parameter dense model, and AFM 3 Core Advanced, a 20-billion-parameter sparsely activated multimodal model optimized for Apple's most capable silicon—and three server-based models running on Private Cloud Compute to ensure user data privacy. These models power an enhanced Siri, advanced photo-editing tools, Image Playground updates, and more expressive voices. Apple highlights their innovative architecture, particularly AFM 3 Core Advanced's novel sparsely activated design using Instruction-Following Pruning that stores most parameters in flash memory while activating only needed ones in DRAM. Evaluation results show significant improvements over previous generations, with AFM 3 Core preferred over its predecessor on 45.6% of prompts, and AFM 3 Cloud on 64.7% of text-based queries.

21h, 62p, 82 comments

Cannibalism

Jack Dorsey, Sebastian Siemiatkowski and Matt Biilmann are among the CEOs racing to hype AI as the industry confronts its own disruption and fears a SaaS‑pocalypse. Brett Kosinski notes that even state‑of‑the‑art models achieve only about 90 % accuracy in summarizing notes, with a 10 % error rate that can misattribute speech and inflate claims, lacking an objective oracle. The article warns of a race to the bottom on price, a potential cambrian explosion of low‑quality content, wage pressure and wealth concentration by investors, while regulation and labour could be the only counterbalance now. The bitter irony is that tech, once the disruptor, may be cannibalising itself.

8h, 424p, 374 comments

Siri AI

Image

Apple Intelligence and Siri integrate AI seamlessly across devices, enhancing personalization, privacy, and functionality through photo editing, accessibility, and cross-platform collaboration, ensuring robust performance.

4h, 65p, 6 comments

Looking Forward to Postgres 19: Query Hints

Postgres 19 introduces core-supported query advice modules (pg_plan_advice and pg_stash_advice), developed by Robert Haas, to address long-standing criticisms of query hints. Instead of embedding hints in SQL (which caused maintenance issues), advice is applied via GUC parameters or stored stashes keyed by query ID, keeping queries clean. The advice constrains the planner's search space without overriding its decisions, degrading gracefully if inappropriate (marking nodes as disabled). Users can specify scan methods (e.g., INDEX_SCAN), join orders (using JOIN_ORDER syntax), join methods, and parallel execution controls. The system provides detailed feedback in EXPLAIN outputs, showing which advice was matched or failed. Stashes offer production-grade, configurable, persistent tuning without app code changes. This represents a decade-long resolution to the debate, offering a controlled, scalable escape hatch for optimizer failures.

6h, 157p, 126 comments

Show HN: Gitdot – a better GitHub. Open-source, written in Rust

The main content highlights an upcoming Build Something Great initiative for Week 20 running from May 24–31, 2026, and announces the v0.2 release focused on infrastructure and issues with an ETA of July 15, 2026. The page also lists trending projects and repositories under categories like gitdot, bkdevs, and truce-audio, alongside a reference to a homebrew-gitdot project and a truce vaibhav135 initiative.

11h, 102p, 44 comments

Launch HN: Intuned (YC S22) – Build and run reliable browser automations as code

Image

Summary: Intuned is a code-first browser automation platform using an AI agent to generate, deploy, and maintain production-ready Playwright scrapers, crawlers, and RPA workflows. Key features include built-in stealth mode, authentication handling, scheduling, auto-scaling, and AI-driven self-healing when sites change. Users can build automations via prompts or code templates in TypeScript/Python, deploy them via CLI, SDKs, or API, and monitor execution through integrated logs, session recordings, and metrics. The platform emphasizes low-code AI editing (e.g., "Fix with AI" command) and offers managed automation services for large-scale deployments.

19h, 175p, 148 comments

1worldflag: A blue dot on a transparent background

Image

TheOne World Flag project promotes global unity through a blue sphere or dot symbolizing Earth, aiming to transcend national divisions by emphasizing shared humanity. It features workshops like "Flag Your ID" in Ukraine, a Nomad magazine ambassador program with handcrafted flags, and global exhibitions, encouraging solidarity via a universal symbol without replacing existing flags.

4h, 11p, 8 comments

Apple Watch for Your Kids

Image

Apple introduces the Apple Watch For Your Kids, enabling parents to pair a GPS + Cellular Apple Watch with their iPhone to provide children with communication, safety, and activity features without needing their own iPhone. Kids can make calls, send texts, and use voice commands with approved contacts, while parents can track their location via Find My, set Schooltime restrictions during school hours, and receive alerts for arrivals or departures. The watch includes durable, water-resistant designs (up to 50 meters), kid-friendly workouts like soccer and dance, and customizable watch faces. Features like Emergency SOS and Activity Awards promote safety and motivation, while Apple Music and the App Store offer entertainment and learning options. The Apple Watch SE 3 is highlighted with adjustable bands and 5G connectivity for faster performance, priced at $199 with cellular. Setup is managed through the Watch app on the parent’s iPhone, and financing options are available. Not all features may be available depending on configuration, and water resistance ratings apply specific usage guidelines. Security and service limitations are noted for Emergency SOS and Apple Cash for users under 18.

13h, 212p, 99 comments

Zig by Example

Image

The webpage discusses a GitHub project focused on education and open-source development using Zig programming language. It highlights resources for teaching Zig through practical examples, emphasizes features like security and developer tools, and points to a community-driven platform for learning and collaboration. The content is geared toward developers at different company sizes and focuses on modern development practices.

5h, 7p, 0 comments

Good Internet magazine on indefinite hiatus

Good Internet, a digital magazine founded by Alexandra, announced a server breach in late May 2026 that redirected links to malicious downloads but did not compromise personal data. The founder, citing burnout, financial losses, and life changes—including an upcoming marriage and legal issues—has decided to pause subscriptions and operations indefinitely. While print editions of the Spring 2025 issue remain available, all 2026 subscriptions will be canceled, with refunds offered. The website will stay active as an archive through 2025, with potential revival considered in 2027. Alexandra expressed gratitude to supporters and invited contributions to sustain the project during this transition.

5h, 14p, 4 comments

Show HN: Courtside – TUI for NBA Games

Image

Courtside is a terminal UI application written in Go that allows users to view NBA games and scores directly in their command line interface. The project can be installed either from source using git clone and `go build`, or with the `go install` command. Once installed, users simply run `courtside` to navigate through live and scheduled NBA games, with features displaying game lists, detailed game information, and league standings. The application consists of backend, UI, and image components to provide a complete terminal-based NBA viewing experience.

1h, 11p, 1 comment

Berkshirehathaway.com – The Perfect Minimalist Website

Warren Buffett addresses fraudulent claims regarding his endorsements of investment products and political candidates, emphasizing Berkshire Hathaway's focus on transparency. The page outlines key sections including annual reports, SEC filings, corporate governance, and letters from Buffett and Charlie Munger to shareholders. It also highlights Berkshire Hathaway Energy and links to operating companies, underscoring the conglomerate's diversified business interests.

5h, 5p, 0 comments

The sample efficiency black hole

The article emphasizes human sample efficiency's dominance, highlighting AI challenges in data scarcity and contextual adaptability, preserving expertise vital for roles requiring nuanced, dynamic competencies.

16h, 68p, 21 comments

Config Files That Run Code: Supply Chain Security Blindspot

Image

**Summary** A new supply chain security threat, dubbed Miasma, exploits editor and package manager configuration files (e.g., `.cursor/rules`, `.vscode/tasks.json`, `composer.json`, `Gemfile`) to auto-run malicious payloads like Epsilon Stealer. These files leverage hooks triggered by folder opens, IDE agents (Claude, Gemini, Cursor), or dependency installs, bypassing security prompts via one-time trust approvals or headless execution. The worm’s payload, a 4.3 MB encrypted stealer, exfiltrates credentials from tools like AWS, GitHub, and npm. Attackers camouflage the dropper in GitHub history, making detection difficult, and pipeline-specific triggers (e.g., `npm test`) enable passive compromise. Organizations must audit config files as code, not just dependencies, and prioritize tools with auto-execution logic as critical attack surfaces. Tags: Malware, GitHub, SupplyChainSecurity, AIAgents, ModernWebThreats

4h, 8p, 1 comment

FP8 Is All You Need (Part 1): Debunking Hardware FP64 as the HPC Holy Grail

Image

The paperchallenges the conventional belief that native hardware FP64 is essential for HPC, demonstrating that on AI‑optimized GPUs such as the B300, using FP8 combined with the Ozaki Scheme II can recover full FP64 accuracy and deliver dramatically higher performance—up to ~500 TFLOPS on B300 and surpassing even H100 baselines—supported by a Tensor‑Memory Equilibrium model and register‑level fusion techniques identified by Satoshi Matsuoka.

18h, 78p, 49 comments

GitHub Is Down

Image

GitHub reported a disruption affecting Issues and Pull Requests on June 8, 2026, with degraded performance and availability first noted at 07:11 UTC and subsequent updates through 07:32 UTC indicating an ongoing investigation.

2h, 8p, 2 comments

macOS 27 requires Apple Silicon, as Apple draws down the Intel Mac era

Image

Apple’s upcoming macOS 27 Golden Gate will exclusively support Macs with Apple Silicon chips, marking the end of Intel Mac compatibility, according to the company’s WWDC 2026 announcement. Intel-based Macs running macOS 26 Tahoe will receive security and Safari updates for roughly two more years, while those on macOS 15 Sequoia get one additional year. Though Rosetta 2 will still enable Intel app compatibility in macOS 27, Apple plans to gradually restrict the technology, focusing on legacy gaming. The transition mirrors past shifts, such as the PowerPC deprecation, as Apple strips Intel code from the OS. Even some Apple Silicon Macs may lack advanced Apple Intelligence features, which require M3 chips and 12GB RAM. The first developer beta is available now, with a public release set for July and the final version slated for fall.

8h, 53p, 17 comments

The IRS Moved IT and HR Staff to Process Taxes. It's Not Going Well

Image

The IRS reassigned more than 1,000 IT and HR staff with no experience in taxpayer services to process tax returns this year, following workforce cuts that reduced its staff by roughly 25,000 employees since President Donald Trump took office. The employees underwent nine weeks of intensive training, but at least several hundred failed their initial certification exams, with some facing potential disciplinary action if they do not improve. Staff described the training as overwhelming and demeaning, with one employee calling it a "nightmare," and many expressing uncertainty about handling even basic tax forms. The IRS emphasized that reassessments and additional coaching would support employees, but officials also acknowledged the challenging year and the possibility of extending the reassignments indefinitely. Critics speculate the moves may become permanent, effectively pressuring employees to leave without formal layoffs, as the agency grapples with a multiyear backlog and reduced capacity ahead of tax season.

7h, 40p, 5 comments

Man jailed for a month despite Flock showing he was 5 miles from crime scene

Image

An innocent man, Hugo Parra, was jailed for a month by San Diego police despite Flock camera data showing he was five miles from a crime scene during an alleged carjacking. Police relied on flawed Flock alerts, superficial identifications (like a hoodie color and beard), and a witness’s vague description to arrest him, ignoring Parra’s cellphone location and car details that contradicted their claims. His attorney argues the technology misuse and failure to vet evidence led to wrongful detention, sparking a lawsuit seeking $1.5 million in damages. The case underscores risks of overreliance on surveillance systems, raising concerns about privacy abuses and systemic errors in law enforcement practices.

7h, 37p, 2 comments

For the 2nd time in weeks, Microsoft packages laced with credential stealer

Image

Microsoft packages compromised for the second time in weeks with a credential-stealing malware called Miasma that activates when opened in AI coding agents. The 73 malicious packages bypassed security by using legitimate Microsoft OIDC tokens to gain access to cloud environments including AWS, Azure, and GCP. The attack, linked to threat actor TeamPCP, exploits the trust model of modern engineering ecosystems by stealing maintainer credentials and publishing packages with valid SLSA provenance. The same Microsoft account was compromised in both May and recent attacks, raising concerns about inadequate remediation. The malware generates unique encrypted payloads for each infection, making traditional detection methods ineffective. Anyone who used the affected packages should thoroughly investigate their systems for potential credential theft.

14h, 556p, 413 comments

Anti-social: It's fads, not friends, which now dominate social media feeds

Image

Social media platforms are increasingly dominated by short-form video content and algorithm-driven feeds, shifting away from personal interaction for many users. Younger demographics, in particular, are consuming professionally produced content from influencers rather than engaging with friends’ posts, reflecting a trend toward passive consumption. Only 33% of users post daily, with 49% of social media users in France described as "active only occasionally" and similar declines in the UK and US. Platforms like TikTok and Instagram prioritize AI-curated content to maximize ad revenue, leading to a divide where entertainment and discovery coexist with private messaging apps like WhatsApp for more intimate interactions. Small businesses now face pressure to create engaging, high-quality content to compete, transforming social media into a job requiring roles like content creator and trend spotter. Despite user concerns about superficial interactions and data privacy, ad revenue continues to grow, with global social media ad spending projected to hit $317 billion by 2026. Platforms offer tools to prioritize friend-based content, but most users opt for algorithmically tailored feeds, cementing the shift from social networking to entertainment consumption.

5h, 17p, 8 comments

Apple bets cheaper AI will woo small developers

Image

At its WWDC keynote, Apple said developers with fewer than 2 million first‑time App Store downloads can use its Foundation Models through Private Cloud Compute with no cloud API cost, a move meant to attract indie developers by removing infrastructure fees and to counter the growing expense of AI experimentation.

2h, 9p, 0 comments

The Grate Cheese Robbery

Image

Neal’s Yard faced a major theft in 2024 when 22 tons of artisanal British cheddar, valued at $400,000, vanished after a fragrant $413,793 theft, highlighting organized crime’s shift to targeting premium, labor-intensive cheeses. Artisanal cheese—handmade, perishable, and complex to resell—has become a prime target, unlike bulk “block” cheese, due to its high value and difficulty in re-packaging. The heist mirrors a global trend: cheese thefts, often organized and sophisticated, exploit digital channels and sanctions (like Russia’s food ban post-Ukraine war). Historical cases, from Jamie Montgomery’s 1998 stolen cheddar to recent Russian smuggling operations, underscore how thieves leverage both traditional and modern methods. Cheesemakers now use microchips to combat counterfeiting, but AI-driven fraud poses new challenges. The thefts reflect broader supply chain vulnerabilities, exacerbated by geopolitical tensions and digital anonymity, with little deterrent for criminals facing relatively light penalties.

6h, 47p, 27 comments

Fooling Go's X.509 Certificate Verification

Image

Daniel Mangum details a subtle discrepancy in Go's X.509 certificate verification where a leaf certificate signed by a CA using UTF8String encoding for its Subject/Issuer fields fails validation against a CA root certificate encoded with PrintableString, despite both certificates appearing identical to OpenSSL and verifying successfully there. The root cause lies in Go's crypto/x509 package performing strict byte-for-byte comparison of the raw RawSubject and RawIssuer fields (DER-encoded ASN.1 tags 0x0c vs 0x13) when building certificate chains, whereas OpenSSL treats the differing string types as equivalent. This strict matching leads to a "certificate signed by unknown authority" panic in Go, highlighting a fail-closed behavior that can cause outages if tooling evolves to change string encodings between long-lived CA and short-lived leaf certificate generation.

9h, 235p, 464 comments

Apple WWDC 2026

Image

The webpage details information about an upcoming special event at Apple, including streaming options, device requirements, and access points for attendees. It outlines the event schedule and accessibility features, such as ASL and non-ASL streams, while emphasizing the availability of related products and services. The content highlights technical specifications and usage guidelines for watching the live event.

8h, 35p, 7 comments

Using XDG-Compliant Config Files (2024)

wxWidgets has updated its configuration file handling to align with the XDG Base Directory Specification, addressing long-standing conventions that placed application data in the user’s home directory as dot files. Since wxWidgets 3.3.0, the wxFileConfig class now defaults to storing new configuration files in the XDG-compliant ~/.config directory (or $XDG_CONFIG_HOME) if no existing dot file is present, ensuring backward compatibility by preserving legacy setups. Additionally, developers can migrate existing configuration files to the new location using the new wxFileConfig::MigrateLocalFile() function, which simplifies transitioning older applications without disrupting users. The article notes that apps requiring legacy behavior must explicitly use the wxCONFIG_USE_HOME flag, while emphasizing that these changes streamline adherence to modern Linux standards without additional effort for new installations.

23h, 755p, 385 comments

Dopamine Fracking

Dopamine fracking is a metaphor describing the harmful practice of overexploiting resources—money, data, or social capital—to extract intense, concentrated dopamine hits from activities, often stripping them of complexity and authenticity, similar to fracking oil. The term critiques how online culture prioritizes commodified, synthetic experiences (e.g., viral trends, simplified content) over nuanced, organic engagement, leading to cultural erosion, addiction, and loss of meaning, likened to replacing fresh strawberries with artificial flavor extracts.

6h, 38p, 8 comments

Donut Lab's 'solid-state' battery exposed as regular li-ion in investigation

Image

An investigation by battery researcher Ziroth and over 20 independent experts has revealed that Donut Lab's claimed "miracle" solid-state battery is actually a standard lithium-ion cell. The company raised approximately $25 million from over 1,300 small investors based on false claims of 400 Wh/kg energy density and 100,000-cycle life. Electrochemical evidence, including voltage curves and graphite anode expansion data, conclusively proves the cell used lithium-ion chemistry rather than the promised sodium-ion solid-state technology. The investigation further revealed that the technology was sourced from a German company, CT Coatings, and that claims regarding the shipment of production vehicles were verifiably false. Finnish financial and criminal authorities are reportedly investigating the company for fraud.

21h, 150p, 13 comments

Playing with Vision Embeddings

Image

Ina June 2026 blog post, Preston Jensen explores the 384‑dimensional embeddings of the DINOv3 ViT‑S model (Siméoni et al., 2025), which maps images to a compact vector while remaining agnostic to language. By leveraging the model’s differentiability and its invariance to augmentations, the author generates images that correspond to arbitrary points in this space, revealing that the embeddings capture a rich mixture of visual concepts. Using a Sparse Autoencoder (SAE) trained on DINOv3 CLS embeddings, they decompose embeddings into roughly 12 000 interpretable directions, enabling the isolation of specific features such as a single whole strawberry versus many small strawberries, or a corn‑kernel arch versus a triumphal arch. Visualization of the feature space with UMAP on ImageNet Val shows clustered groups of co‑activating features, illustrating both the power and the opacity of superposition in vision transformers.

1d, 147p, 83 comments

Algorithmic Monocultures in Hiring

Researchers from Stanford, Chapman, and Northeastern universities conducted the largest empirical study of algorithmic hiring, analyzing 3.4 million applications to 156 employers across 11 market sectors. The study reveals significant racial disparities in algorithmic hiring, with 25.87% of applications from Black applicants and 14.74% of applications from Asian applicants directed to positions that adversely impact them according to Title VII standards. These disparities only became apparent through position-by-position analysis, as aggregate data had previously concealed them. The research also demonstrates that algorithmic monocultures—where multiple employers use algorithms from the same vendors—lead to systemic rejections, with 10% of applicants submitting 4 applications being rejected everywhere, significantly exceeding expected rates. The study concludes that data access barriers inhibit independent research and recommends policymakers measure adverse impact per position, strengthen market surveillance, monitor algorithmic monocultures, and expand researcher access to hiring algorithm data.

3h, 9p, 5 comments

Ask HN: How do you cope when your startup contracts?

The article on Hacker News discusses a technical lead's repeated experience of being left as the sole employee in his department during startup contractions, absorbing the workload of laid-off colleagues while facing increased scrutiny and unrealistic job titles. The author, who has worn multiple hats in engineering, product, and support, questions whether this pattern reflects bad luck with startups, if he should have pushed harder for role clarity during layoffs, and whether he should embrace his generalist skills as a deliberate career path despite market challenges. He highlights the tension between his adaptability and the lack of recognition, as well as the job market's difficulty amid economic uncertainty, while seeking advice from others who may have faced similar "Red Mage" roles in startups.

9h, 12p, 3 comments

Good type against all odds

Image

Marcin Wichary examines how typographic quirks stem from display limitations, noting that the familiar italic slant is a side effect of top‑to‑bottom line scanning (akin to rolling shutter) and could have been reversed, then highlights intentional pixel‑level crafts—such as small caps, a train carriage icon, a stadium icon, vertical ligatures, inventive underlines, and hyphenation/easter‑egg finds—that reveal delightful details when designers embrace constraints.

1d, 479p, 227 comments

How's Linear so fast? A technical breakdown

Image

Linear achieves exceptional speed by implementing a local-first architecture that minimizes network latency. The app stores a database in the browser using IndexedDB, applying mutations locally before syncing to the server asynchronously. This eliminates loading states for most operations. Their optimization strategy includes aggressive code splitting (reducing shipped code by 50%), parallel preloading of JavaScript chunks, and a service worker that caches assets in the background. The sync engine uses granular MobX observables so updates affect only specific components rather than entire views. Design choices prioritize keyboard shortcuts over mouse interactions and animate only GPU-friendly properties like transform and opacity. Unlike traditional CRUD apps that take ~300ms for updates, Linear completes the same operations in milliseconds by treating the browser as the primary database and the server as a sync target.

Image
Image