Hacker News

The Front Page

Vol. I June 16, 2026 Price: A cup of coffee

1h, 167p, 76 comments

Running local models is good now

Vicki Boykis reports that local LLMs have reached a point where they are “surprisingly good,” citing his experience on a 2022 M2 Mac with Gemma‑4‑26B‑A4B and the newer Gemma‑4‑12B‑QAT. Using tools such as llama.cpp, LM Studio, and Ollama, he demonstrates agentic coding tasks—including refactoring Python notebooks, generating type‑hints, proofreading, writing tests, and bootstrapping a recommendation model—at about 75 % of frontier‑model speed and accuracy. He details a Docker‑based setup with the Pi agent harness, custom models.json, and a compose configuration that isolates execution. While acknowledging remaining limits (inference speed, context size, prompt mismatches), Boykis emphasizes the rapid tooling improvements, introspection capabilities, and the growing viability of local models for development work, though he stops short of declaring them production‑ready.

5h, 173p, 58 comments

Correlated randomness in Slay the Spire 2

Correlated randomness (CRNG) in Slay the Spire 2 causes multiple in‑game RNG streams to be predictably linked because each is seeded with the same base seed plus a fixed offset, and Unity’s System.Random produces outputs linear in that seed. This leads to exploitable biases such as Neow’s Bones overwhelmingly yielding the Debt curse in Underdocks, the impossibility of receiving Rebound from the Trash Heap event, and dramatically different first‑fight potion drop rates (≈76% in Underdocks vs ≈4% in Overgrowth). The author details how these correlations affect relic rarity, card transforms, orb targeting, event outcomes, and gold rewards, and provides extensive probability tables. They argue the issue is a bug, easy to fix by replacing System.Random with a non‑linear PRNG (e.g., PCG32) or using counter‑based generators, and urge Mega Crit to address it before the game leaves early access.

19h, 1450p, 274 comments

A backdoor in a LinkedIn job offer

Roman Imankulov recounts a LinkedIn‑based supply‑chain attack where a recruiter from a crypto startup sent him a public GitHub repo and urged him to run npm install. The repository, a React frontend with Node backend, contained a hidden backdoor in app/test/index.js that assembled a URL (https://rest‑icon‑handler.store/icons/77) and executed any code returned by that server. The backdoor was triggered automatically via the prepare script in package.json, which runs on install, allowing the malicious payload to run without user interaction. The commit history was fraudulently attributed to a real developer, and the recruiter’s LinkedIn profile used the identity of a well‑known arts journalist, both borrowed to lend credibility. Imankulov’s use of a read‑only analysis agent exposed the code quickly; he reported the repo to GitHub and the recruiter to LinkedIn, warning that such deceptive job offers can compromise anyone who follows the instructions without caution.

7h, 108p, 17 comments

Getting Creative with Perlin Noise Fields

Manohar Vanga documents a personal experiment in which he created 25 distinct generative artworks using Perlin noise flow fields in Processing. By iteratively tweaking parameters such as background color, particle opacity, line thickness, stroke caps, color palettes (including a golden‑ratio HSB generator), shapes (lines, arcs, circles, rectangles, triangles), and noise scale, he produced a diverse series ranging from monochrome “charcoal‑pencil” textures to colorful nebula‑like compositions and leaf‑inspired autumn scenes. The project highlights how self‑imposed constraints and relentless iteration can foster creativity, turning simple algorithms into increasingly complex visual results.

17h, 515p, 302 comments

Banned Book Library in a Wi-Fi Smart Light Bulb

Rick Osgood describes how he turned a Wi‑Fi smart light bulb into a covert “Banned Book Library” by flashing the bulb’s ESP32‑C3 with custom firmware that creates an open Wi‑Fi access point and a web server hosting a small collection of censored e‑books. He details the hardware teardown, challenges of accessing ESP32 pins, attempts to add microSD storage, and ultimately reshaping the ESP32 partition table to enlarge the SPIFFS partition for up to 2 MB of book files. Using Arduino and ESP‑IDF, he built a minimal firmware with a captive‑portal, admin page for LED color matching, and OTA update support via a custom safeboot partition. He notes the 4 MB flash limit restricts the library to a few books, making each device a personal “dead drop,” and suggests future enhancements such as color‑control sliders and mesh networking.

8h, 82p, 26 comments

Understanding the rationale behind a rule when trying to circumvent it

The article discusses the importance of implementing best practices for process and thread callbacks in .NET development, emphasizing the need for quick, non-blocking operations to maintain system performance. It highlights common pitfalls where long-running functions can cause delays, urging developers to use asynchronous tasks like System Worker Threads instead of relying on callbacks to prevent blocking behavior. The guidance aims to ensure efficient UI responsiveness and avoid frustrating end users caused by sluggish system responses.

19h, 285p, 155 comments

I Love the Computer

Michael Enger reflects on his lifelong passion for computers, tracing it from a childhood IBM 486 in Norway to teenage exploration of the early Internet in Malaysia, through self‑taught programming, university studies, and a career in web and audio development. He mourns the current AI hype and the commercialization that has corrupted the once open, community‑driven tech culture, yet notes a resurgence of accessible, decentralized tools. Despite the industry’s shift toward profit‑driven “brogrammers” and restrictive platforms, Enger remains devoted to the computer as a source of personal identity and creative freedom.

17h, 203p, 91 comments

Why I email complete strangers

Zachary Kai reflects on why he repeatedly emails strangers despite the anxiety it provokes, arguing that email’s longevity and intentionality make it a unique, meaningful medium. He traces email’s history to Ray Tomlinson’s 1971 invention, cites Lindy’s law to explain its endurance, and contrasts its thoughtful pace with the instant pressure of social media. Kai shares practical guidelines for reaching out—research the recipient, be specific, keep messages brief yet courteous, use engaging subject lines, and avoid overtly commercial requests—and lists resources like Derek Sivers’ open inbox and the IndieWeb community for finding contacts. He concludes that while the worst outcome is silence, the potential for genuine connection makes the risk worthwhile.

19h, 240p, 186 comments

What job interviews taught me about Kubernetes

What job interviews taught me about Kubernetes – The author, after interviewing at numerous companies, found that almost every organization now uses Kubernetes, a shift from five years ago when only a few adopters existed alongside VM/systemd and serverless setups. CTOs cite three main non‑technical reasons: uniformity – a single deployment method for all services; standardized, hireable knowledge – Kubernetes serves as a common language that makes onboarding and on‑call work easier; and traceability – GitOps workflows (Helm, FluxCD/ArgoCD) provide audit trails and compliance benefits. While the author acknowledges Kubernetes’ complexity and suggests it may be overkill for very early‑stage teams, they note that managed services and tools like Helm have lowered barriers, making Kubernetes the default choice once a second engineer joins and knowledge retention becomes critical.

22h, 226p, 77 comments

Game Engine White Papers: Commander Keen

Game Engine White Papers: Commander Keen is a 214‑page, full‑color book released on March 28, 2026, detailing the development of the original Commander Keen game from 1990. Available in print on Amazon and as a free high‑resolution PDF, the book covers the era’s hardware (80286 CPUs, EGA video, sound cards, keyboards), game assets, engine design, and the creation of the CGA version. The author spent three years writing it, provides the source code on GitHub, and invites readers to report any inaccuracies via GitHub issues.

3h, 22p, 29 comments

AskHN:How do you handle skill atrophy from using coding agents?

AskHN discussion on Hacker News explores how developers can prevent skill atrophy as they increasingly rely on AI coding agents. One commenter argues that using AI exclusively makes skill retention impossible, while others suggest focusing on higher‑level tasks—such as architecture decisions, requirements analysis, technology trade‑offs, and cross‑system debugging—that AI cannot yet replace, or cultivating new capabilities altogether. A perspective also notes that guiding AI still requires cognitive effort, recommending simple brain‑exercise like crossword puzzles for those idle.

1h, 3p, 1 comment

Adobe Alternatives: list of alternatives for Adobe software

Adobe‑Alternatives is a community‑maintained GitHub repository that compiles free, open‑source, and low‑cost replacements for the full range of Adobe Creative Cloud apps. The README lists alternatives for Photoshop, Illustrator, Animate, InDesign, Substance, Lightroom, XD, Premiere, Acrobat, Bridge, Dreamweaver, ColdFusion, Audition, After Effects, Mixamo and related tools, categorizing each by license type (open‑source ✨, free ⭐️, paid 💵, subscription 🔒, trial ⌛). Options include well‑known programs such as GIMP, Krita, Inkscape, Blender, DaVinci Resolve, Figma, Handbrake, and many others, with platform compatibility details and icons indicating cost or trial status. The repo invites contributions via issues or pull requests to keep the list current.

1d, 142p, 49 comments

Making glass-to-metal seals for home­made vacuum tubes

Maurycy details practical methods for creating reliable glass‑to‑metal seals in homemade vacuum tubes, focusing on borosilicate glass. He explains why common metals like copper fail due to thermal‑expansion mismatch, and shows that tungsten wire (up to ~0.7 mm) or copper foil/disc seals can succeed, though copper requires careful heating and shaping. He describes plating steel with copper, using borax treatment, and testing leaks with a capacitive glow discharge. The post also notes limitations of soda‑lime glass, alternative seal concepts (houskeeper and knife‑edge seals), and practical tips such as annealing and handling ultra‑thin tungsten.

1d, 516p, 706 comments

Hetzner Price Adjustment

Hetzner announced a price adjustment effective 15 June 2026 08:00 CEST for new dedicated‑server orders, server rescales and cloud instances. Existing orders placed before that date retain old rates, but delivery after the change uses the new pricing. The update lists revised monthly and setup fees (excluding VAT) for a full range of dedicated‑server models across Falkenstein, Helsinki, Nuremberg and other sites, as well as new hourly and monthly rates for cloud products (CAX, CCX, CPX, CX) in Europe, the US and Singapore, with most prices increasing. Limited‑offer servers are available only while stock lasts, and customers are directed to the server auction for additional discounted options.

14h, 55p, 9 comments

The 90-year-old idea behind JEPA models: Canonical Correlation Analysis

Shon Czinner explains that modern Joint Embedding Predictive Architecture (JEPA) models trace back to Harold Hotelling’s 1936 Canonical Correlation Analysis (CCA), which seeks linear transformations that maximize correlation between two data matrices. He shows that JEPA’s objective—minimizing the mean‑squared error between predicted and target embeddings—mirrors CCA’s formulation, differing mainly in the absence of whitening constraints that cause collapse without additional regularisation such as isotropic Gaussian enforcement (e.g., SIGReg). The post argues that while JEPA adds non‑linear encoders and predictors to the classic CCA idea, the core goal remains the same: find transformations yielding maximal multidimensional correlation, and credit for practical breakthroughs lies in preventing collapse and scaling the method. References span from Hotelling’s original paper to recent “Deep CCA” and JEPA literature.

6h, 143p, 91 comments

Commodore Releases Flip Phone

Commodore unveils the Callback Starlight Edition, a flip‑phone that blends minimalist design with essential smart features like WhatsApp, Signal, music, maps and a quality camera, while blocking social media, browsers and other anxiety‑inducing apps at the system level. Powered by a custom Sailfish OS built by former Nokia engineers, the device runs select Android apps without Android, emphasizes data privacy and “friendly” technology, and even includes a Commodore 64 emulator with classic games and SID‑chip ringtones. The CEO positions the phone as a purposeful “bridge” between dumb phones and intrusive smartphones, targeting users seeking digital minimalism, reduced FOMO, and protection from tracking and addictive behavior. Pre‑orders are announced with a sign‑up link.

11h, 21p, 4 comments

Honeypot Design

Bruce Ediger shares his extensive experience running various honeypots—WordPress, endlessh, and cowrie—and argues that anyone with the capability should deploy them to waste attackers’ time and deter low‑level abuse as well as large‑scale scrapers. He outlines design considerations, balancing minimal resource use, maximal attacker resource consumption, realistic emulation, extensive logging, and avoidance of collateral damage, noting that trade‑offs must be resolved through practice. Ediger also lists scarce literature on honeypots and emphasizes the ongoing, arms‑race nature of maintaining effective honey‑pot systems.

15h, 116p, 126 comments

Amazon Announces Multibillion-Dollar Data Center in Missouri

Amazon Web Services announced a multibillion‑dollar data‑center campus in Montgomery County, Missouri, projected to create over 400 permanent jobs and thousands of construction roles. The site will support AWS cloud and AI workloads, feature carbon‑free energy from a 138 MW project, and employ advanced water‑conservation methods that use less than 0.1% of the local aquifer. Amazon pledged more than $7 million for community projects, including emergency services, a fairground gathering space, and infrastructure upgrades, while expecting hundreds of millions of dollars in new tax revenue. Governor Mike Kehoe welcomed the investment as a boost for regional infrastructure and economic growth.

22h, 324p, 87 comments

Typst 0.15.0

Typst 0.15.0 introduces variable‑font support, native MathML export, experimental bundle export for multi‑file projects, multiple bibliographies per document, and simultaneous targeting of several PDF standards. New syntax changes include forward‑slash‑only file paths and an enhanced within selector. Layout improvements fix list marker alignment, baseline handling, and paragraph grouping in HTML export. The library adds a file‑path type, map/filter for dictionaries, new math functions (asinh, acosh, etc.), spot colors, and expanded symbol sets. Bibliography, footnote, and numbering systems gain richer features, while CLI gains a typst eval subcommand and better diagnostics. Several deprecated elements are removed and many behaviours are tightened for consistency and standards compliance.

17h, 38p, 7 comments

Micro Radar: a tiny open-source flight radar for your desk

Micro‑Radar is a DIY, open‑source desk‑size flight radar built around an ESP32‑C3 board with a 240×240 IPS screen. The repository provides 3‑D printable enclosure files, a step‑by‑step assembly guide (including optional glass lens and epoxy), and firmware that fetches live aircraft data from the OpenSky API (free accounts raise request limits from 400 to 4 000 per day). After flashing the firmware via PlatformIO, the device creates a Wi‑Fi hotspot for initial configuration, then operates on the user’s network, offering a web interface at http://microradar.local to set location, radar radius, display options, and OpenSky credentials. The project is licensed under MIT and was created as a wedding gift for an aviation enthusiast.

1d, 476p, 220 comments

Apple Foundation Models

Claude for Foundation Models is a Swift package that integrates Anthropic’s Claude language models into Apple’s Foundation Models framework, conforming to the LanguageModel protocol so developers can use the same LanguageModelSession API as for on‑device models. The package enables direct calls from an app to the Claude API (billed to the Anthropic account) with support for model selection, effort levels, streaming, structured output, tool calling (client‑side and server‑side), and image input. It requires iOS/macOS/visionOS/watchOS 27 beta, Xcode 27 beta, and an Anthropic API key, with production apps recommended to route requests through a proxy to hide the key. The documentation covers installation via SwiftPM, quick‑start code, model capabilities, authentication methods, error handling, and limitations such as unavailable prompt‑caching controls. The package is beta‑stage, Apache 2.0 licensed, and not a general‑purpose Messages API client.

4h, 421p, 74 comments

Mechanical Watch (2022)

Mechanical Watch – Bartosz Ciechanowski explains how a traditional mechanical watch works, detailing each component of the movement from the mainspring and barrel that store energy, through the gear train that amplifies rotation, to the escapement and balance wheel that regulate timing. The article covers the click mechanism that prevents the spring from unwinding, the keyless works allowing winding, time‑set, and date‑set functions via the crown, and the automatic winding system that harvests wearer motion. It also describes the assembly of the mainplate, balance, and motion works, and notes the watch’s limitations versus quartz and smart watches while highlighting its engineering artistry.

11h, 421p, 133 comments

The time the x86 emulator team found code so bad they fixed it during emulation

Raymond Chen recounts a story from the early days of Windows when the x86‑32 emulator used binary translation to JIT‑compile guest code. A program that allocated a 64 KB stack buffer performed a stack probe and then initialized the memory. Instead of generating a simple loop, the compiler unrolled the initialization into 65,536 separate byte‑write instructions, producing a 256 KB function. The emulator team found the bloated code “so bad” that they added a special translation‑time check to replace it with an efficient loop during emulation.

2h, 28p, 28 comments

4× RTX Pro 6000 Blackwell on Water, and the One Card That Wouldn't Behave

Four RTX PRO 6000 Blackwell GPUs were water‑cooled to handle a 2.4 kW sustained load for multi‑day AI model training. The pilot card failed after a week due to a cracked VRM power inductor that detached during cold‑pad removal, causing Xid 79/DPC containment errors. The fix was to warm the GPU to ~90 °C before removing pads, then microsolder the inductor at a local repair shop. Applying this procedure to the remaining three cards eliminated failures; the fully water‑cooled rig now runs all four GPUs at full 600 W boost with stable temperatures (~58 °C) and delivers 41 k tokens/s (≈838 TFLOPS) for training and over 8 k tokens/s for inference without throttling.

1h, 36p, 7 comments

Show HN: Hackers for Granny (defense against industrialized elder fraud)

HACKERS FOR GRANNY is a manifesto urging cybersecurity professionals to combat the $64 billion annual “industrialized elder fraud” syndicates operating from Southeast Asian compounds. It outlines how scammers use voice‑cloned calls, deep‑fake videos, and remote‑access tools (Quick Assist, AnyDesk, TeamViewer) to exploit seniors’ loneliness and cognitive decline, overwhelming traditional advice. The piece calls for volunteers to develop and deploy lightweight detection tools—exemplified by the free‑trial “Granny Kate™” app that monitors for five known attack patterns—to gather forensic evidence and protect vulnerable elders, emphasizing community‑driven, non‑violent defense and highlighting recent FTC and DOJ actions against such scams.

2h, 10p, 4 comments

The history of butterfly swimming

The butterfly stroke originated in the 1930s as a faster variant of breaststroke, with American swimmer Henry Myers first using the full butterfly arm motion in 1933 and coach David Armbruster adding the dolphin kick in 1935. Although initially illegal under FINA’s breaststroke rules, the technique became a separate stroke in 1952 and debuted at the 1956 Melbourne Olympics (men’s 200 m and women’s 100 m). The 100 m and 200 m events remain Olympic fixtures, while a 50 m race is contested at World and continental meets. English butterfly laurels include Philip Hubble (silver, 1980), Andy Jameson (bronze, 1988) and Steve Parry (bronze, 2004), with James Hickman holding world short‑course records in the late 1990s.

22h, 165p, 22 comments

How TimescaleDB compresses time-series data

TimescaleDB uses a hybrid row‑columnar engine called hypercore to compress time‑series data up to 98 % by exploiting cross‑row patterns with algorithms such as delta encoding, delta‑of‑delta, Gorilla XOR, run‑length and dictionary compression. Unlike PostgreSQL’s TOAST, which compresses large individual values, hypercore groups rows into batches of ~1,000, stores columns as compressed arrays, and applies column‑specific techniques, achieving typical ratios of 10‑100× for timestamps and floats and 5‑10× for repetitive text. Proper segmentby (grouping column) and orderby (sorting column) settings are crucial; they enable metadata‑driven batch skipping and optimal delta encoding, dramatically speeding range scans and aggregations while reducing I/O 10‑20×. Point lookups and updates on compressed chunks are slower, and high‑cardinality segmentby values can hurt compression. Real‑world tests on IoT sensor tables show > 40× size reduction and > 10× query speedup. The article also provides SQL commands for enabling compression, setting policies, and verifying results.

21h, 57p, 62 comments

Launch HN: Drafted (YC P26) – Models for residential architecture

Drafted, a Y Combinator‑backed startup, is developing AI models that generate complete residential floor plans and exterior elevations from simple design constraints such as square footage, lot shape, and room placement. The system produces 2D and 3D visualizations, allows instant iteration, furniture placement, material swaps, and exports CAD or PDF files for pre‑construction. In its first month, over 120,000 users created 325,000 home designs, aiming to make custom home design faster, cheaper, and accessible to anyone.

19h, 70p, 55 comments

Show HN: Veterinarian turned founder, AI lawn diagnosis

GrassDx offers a free AI-powered service that diagnoses lawn problems from up to four uploaded photos and a ZIP code, delivering an instant, region‑specific treatment plan with options for DIY Amazon products, subscription services, or professional quotes. It claims to identify over 47 conditions, such as fungal diseases, drought stress, and nutrient deficiencies, and provides user testimonials highlighting accurate diagnoses and cost savings. The platform emphasizes no account needed, quick results, and localized advice tailored to grass type and season.

23h, 139p, 201 comments

How memory safety CVEs differ between Rust and C/C++

How memory‑safety CVEs differ between Rust and C/C++ – The author explains that in C/C++ libraries, bugs like passing NULL to curl_getenv cause undefined behaviour but are treated as user‑code misuse, not a CVE for the library, because the language lacks a safe/unsafe distinction and APIs cannot fully express contracts. In Rust, any memory‑safety crash that occurs without unsafe in user code is a soundness bug in the library itself, so a similar misuse would be reported as a CVE. Consequently, Rust CVEs appear “stricter” and raw CVE counts are misleading when comparing the languages, since Rust’s type system guarantees that safe APIs cannot be misused, while C/C++ code can always be misused inadvertently. The post uses libcurl vs. hyper as examples and concludes that comparing CVE numbers without accounting for these semantic differences is unfair.

21h, 96p, 4 comments

Factoring "short-sleeve" RSA keys with polynomials

The Trail of Bits researchers discovered that RSA private keys whose binary representation is unusually sparse in zero‑bits—so‑called “short‑sleeve” keys—can be factored quickly using a polynomial‑based technique. By scanning millions of public keys from Certificate Transparency logs, TLS/SSH scans and other sources, they identified hundreds of vulnerable keys, including 603 RSA and 74 DSA keys generated by old versions of EnterpriseDT’s CompleteFTP software. The flaw stemmed from a type‑mismatch bug in the big‑integer random‑bit generator, which left regularly spaced blocks of zeros in each limb of the key. Exploiting this structure, the team expressed the modulus as a low‑coefficient polynomial, factored it, and recovered the private primes. The vulnerability affected CompleteFTP versions 10.0.0–12.0.0 (RSA) and 10.0.0–23.0.4 (DSA) and was mitigated after the code was refactored to use standard .NET crypto APIs; EnterpriseDT released updates and tools to detect and replace weak keys. The work also highlights the broader risk of similar implementation errors and the power of tailored cryptanalysis.

1h, 3p, 0 comments

A Joy That Is Being Lost

Dave Griffith reflects on his 48‑year programming career, beginning with an Apple II in 1977, and mourns the loss of the “craft” of hand‑coding as AI‑driven “agentic” tools now generate high‑quality software with simple English prompts. While he embraces the productivity gains and believes the shift is positive for humanity, he misses the precise, tactile satisfaction of holding an entire machine in his mind and hearing the “click” when code finally runs, a personal joy he feels is disappearing for veteran programmers. The post, assisted by Claude Opus 4.8, urges future coders to appreciate both the new capabilities and the lost pleasure of meticulous craftsmanship.

6h, 32p, 11 comments

Dutch children are unusually happy and healthy. Because of this walking ritual?

Avondvierdaagse, a four‑evening walking festival, sees half a million Dutch children and families traverse 5‑10 km routes across neighborhoods each summer, rain or shine, earning medals and sweets for completing all four nights. Originating from a 1909 military march, the event now fosters physical activity, resilience and strong peer connections, contributing to the nation’s reputation for happy, healthy youngsters. Organized by volunteers and the Royal Dutch Walking Association, it reinforces Dutch “gezelligheid” – community togetherness – though concerns persist about inclusivity and volunteer sustainability.

15h, 73p, 69 comments

Reviews have become expensive, rewrites have become cheap

Reviews have become expensive, rewrites have become cheap – The author argues that large language models (LLMs) generate fully fleshed‑out code rather than opting for minimal library imports, leading to technically correct but over‑engineered solutions. Consequently, code reviews now consume more time and effort as reviewers must assess and often push back on unnecessary complexity. In contrast, simplifying or rewriting this code with the same AI is fast and cheap, shifting the workflow toward upfront planning and scope definition to prevent bloat, then using AI to refactor excess code after implementation. This change makes the review stage the costlier step, while iterative AI‑driven rewrites become economically attractive.

1d, 137p, 277 comments

Can Europe train a frontier AI model on the compute it owns?

EuroMesh is a GitHub repository presenting a sourced model and brief report that investigates whether Europe can assemble a sovereign, frontier‑class AI model using its existing public compute resources—primarily the EuroHPC supercomputers and national AI Factories—rather than waiting for large‑scale (1 GW) datacenters to secure grid power, which averages a 7.6‑year delay. The analysis, built on a three‑layer model of training efficiency, compute availability, and regional feasibility, concludes that federated low‑communication training can deliver a frontier‑class model by around 2028, effectively outpacing a new gigawatt campus projected for 2033. The repo includes the full report (PDF/Markdown), model specifications, code to regenerate results, data sources on grid‑connection lead times and EU compute inventory, and extensive caveats noting uncertainties in grid estimates, shared compute access, and the unproven nature of large‑scale distributed training.

11h, 114p, 61 comments

OpenAI Losses Increased Nearly 8X in 2025, with Spending Hitting $34B

OpenAI’s audited financials reveal a dramatic surge in losses, rising from a $5.09 billion net loss in 2024 to an estimated $38.5 billion net loss in 2025 despite revenue growth to $13.07 billion. Costs ballooned to $34 billion, driven by $19.18 billion in research and development and $5.73 billion in sales and marketing, while a conversion to a for‑profit structure added a $41.55 billion fair‑value adjustment. The company received $867 million from SoftBank and $303 million from Microsoft in 2025, but paid Microsoft $10.59 billion for R&D and $17.2 billion total in expenses, leaving it with over $50 billion in assets, half in cash, and substantial liabilities to Microsoft. The report underscores concerns about OpenAI’s sustainability and profitability.

1h, 36p, 16 comments

Subquadratic – Introducing SubQ 1.1 Small

SubQ 1.1 Small, the newest Subquadratic Sparse Attention (SSA) model from Subquadratic Inc., offers near‑perfect long‑context retrieval up to 12 million tokens and reduces attention compute by up to 1,000× compared with dense attention. Trained on trillion‑token long‑form data, it achieves 64.5× less compute and 56× faster inference than FlashAttention‑2 at 1 M tokens, while scoring competitively on knowledge, coding, and finance benchmarks (e.g., 85.4% GPQA Diamond, 89.7% LiveCodeBench pass@4). Independent testing by Appen validates the results. The model targets enterprise AI tasks requiring whole‑artifact reasoning—such as financial due diligence, legal contract analysis, and full‑repo code understanding—and will be rolled out to design partners soon, with broader releases planned by year‑end.

3h, 48p, 23 comments

An interview with an Apple emoji designer

Ollie Wagner, one of Apple’s original emoji designers, recounts his 2008 internship on the Human Interface team where he helped create over 300 of Apple’s first emoji from a SoftBank spreadsheet, refining each glyph by hand in Photoshop before approval by Steve Jobs. The project, driven by the need to compete in the Japanese market, was treated seriously by Apple but without foresight of its cultural impact. After the internship Wagner joined Apple full‑time, contributed to the iPad’s design, and now runs YAP Studios in Albuquerque. The interview highlights Apple’s meticulous early emoji work and its lasting global recognisability.

4h, 35p, 10 comments

Unicorn – The Ultimate CPU Emulator

Unicorn Engine is a lightweight, open‑source CPU emulator supporting multiple architectures—including ARM, ARM64, x86/x86_64, MIPS, PowerPC, RISC‑V, S390x, SPARC, TriCore and m68k—with a clean C API and bindings for dozens of languages. Built on QEMU and featuring JIT compilation, it runs on Windows, macOS, Linux, Android, BSD and Solaris, and is thread‑safe and GPL‑v2 licensed. Since its 2015 debut at BlackHat USA, Unicorn has seen regular releases, most recently v2.0.1 (Nov 2022), and earned recognition such as Alibaba Cloud’s Asian Star 10x10 award for its impact on cybersecurity and broader fields. The site offers source code, binaries, documentation, and a history of community contributions and updates.

6h, 83p, 11 comments

Trinket.io shutting down, so we saved it and hosted it a trinket.strivemath.org

Trinket is now offered as a free, community‑hosted coding platform by Strive Math at trinket.strivemath.org. Built on the open‑source Trinket project, it provides browser‑based development for languages like Python, HTML, and Java without downloads or paid plans. Users can instantly edit and run code, save and share projects, and educators can create interactive courses and monitor student progress. Existing trinket.io users can import their projects via an export zip. The service is open source on GitHub and operates under a CC0 1.0 license.

11h, 116p, 16 comments

Show HN: Garden of Flowers – an archive of pictorial typography before ASCII art

Heikki’s Garden of Flowers is an extensive digital archive of typographic and decorative artifacts spanning the 16th‑century to the present, curated by Heikki. It includes over 2,500 high‑resolution images of historic type specimens (e.g., Caslon, Louis Luce, French and German foundries), ornamental borders, head‑and‑tail pieces, modular and silhouette styles, printing‑press photographs, and contemporary graphic works. The collection is organized by geography (Germany, France, England, Spain, Finland, Estonia, etc.), period, and subject (flowers, birds, ships, machines, ex‑libris, rule pictures, etc.), and features contributions from notable typographers and designers such as Alvin Lustig, Albert Schiller, Paulo de Cantos, and El Lissitzky. It serves as a resource for scholars, designers, and enthusiasts exploring the evolution of printing type, decorative typography, and graphic design.

1d, 1321p, 408 comments

Iroh 1.0

Iroh 1.0 launches as a stable, open‑source networking library that replaces traditional IP addresses with user‑controlled cryptographic dial keys, enabling devices to be securely addressable anywhere. The release highlights multi‑path QUIC, NAT traversal, local‑first connectivity, WASM support, and custom transports (BLE, LoRa, Wi‑Fi Aware, Tor). Over 200 million endpoints were created on public relays in the past month, and the platform now offers official APIs for Python, Node.js, Swift, and Kotlin alongside the Rust crate. Wire protocol and API stability are guaranteed for v1, with a defined support schedule for future versions and public relay deprecation timelines. Iroh aims to make internet connections more direct, efficient, and secure, inviting developers to integrate it via docs, quickstart guides, and community channels.

20h, 113p, 39 comments

What every coder should know about gamma (2016)

What every coder should know about gamma explains that gamma correction—using a non‑linear encoding (approximately γ = 2.2) to match human perception—allows 8‑bit images to store visual detail efficiently, but it also means that most image‑processing algorithms assume linear light values and will produce incorrect results if applied directly to sRGB‑encoded data. The article covers the physics of light vs. perceived brightness, the gamma transfer functions, the distinction between plain gamma and the sRGB colour space, and the need for proper gamma calibration. It demonstrates common errors in gradients, colour blending, alpha compositing, image resizing, antialiasing, and physically‑based rendering when gamma is mishandled, and provides a short quiz to identify developers likely to encounter these pitfalls. The conclusion urges developers to adopt a linear workflow—decoding to linear space before processing and re‑encoding for display—and to explicitly document colour‑space assumptions.

23h, 105p, 51 comments

Show HN: Fata – Spaced repetition to fight skill rot from AI coding

fata is a free, mobile‑first platform that teaches developers full‑stack fundamentals through bite‑size, science‑backed courses such as SOLID principles in Python, React Core, Rust for Polyglots, and TypeScript Foundations. Users can learn directly in a browser or via iOS/Android apps, with spaced‑repetition AI reviews, offline access, and progress tracking. The service offers three tiers—Core (basic courses and community features), Pro (advanced professional content and career pathways), and Organization (team management, custom assessments, and analytics). Over 10,000 developers have joined, and fata seeks partners in education and business to expand its content and integrations.

Job, 1d

Dalus (YC W25) Is Hiring a Senior Software Engineer in Germany

Dalus is hiring a senior software/frontend engineer for its Germany office (preferably Munich) to lead the UI/UX of its AI‑powered platform that modernises hardware system design for rockets, satellites, EVs and similar complex products. The role demands 5‑8 years of React experience, strong visual design sense, and the ability to own the frontend in a small, autonomous team, with startup experience and knowledge of engineering tools seen as a plus. Compensation is €70‑90 K plus equity, and the position is open to candidates in the DACH region without visa requirements. Applicants will interview with the CEO, CPO and CTO and complete a hands‑on code challenge.

23h, 129p, 56 comments

Boot Naked Linux

Nick Moore outlines how to create a minimalist Linux system that boots in under a second by stripping everything except the kernel and a single init program. He provides a step‑by‑step guide: writing a tiny C init that prints a message and reboots, packing it into a one‑file initrd using cpio, and booting the kernel with QEMU/KVM. Moore explains mounting devtmpfs to access block devices, reading raw disk data, and configuring a USB‑bootable EFI stick with a unified kernel built via ukify. He also details kernel configuration options (using make tinyconfig and menuconfig) to reduce size to 1‑2 MB, enable essential drivers, filesystems, and EFI stub support. The article demonstrates the complete workflow from kernel preparation to real‑hardware boot, highlighting speed, reduced attack surface, and potential for further customization.

1d, 210p, 81 comments

Openrouter Fusion API

OpenRouter’s Fusion is a multi‑model routing service that turns a single prompt into a deliberation among several expert models, optionally adding web search and fetch. The panel of models (default “Quality” preset or a cheaper “Budget” option) runs in parallel, after which a judge model synthesizes their outputs into a structured analysis highlighting consensus, contradictions, gaps and unique insights, then generates the final answer. Pricing is the combined cost of all panel members plus the judge call. Users can customize the panel and judge via the analysis_models and model fields, view activity logs, and access API documentation that is compatible with OpenAI‑style endpoints.

1d, 187p, 104 comments

Teenagers Stayed Overnight at Their School and Found Hidden Ancient Roman Ruins

Students at Liceo Cavour, a high school across from the Colosseum in Rome, uncovered a hidden mid-second-century C.E. Roman villa after occupying the building in January 2021 to protest expanded remote learning during the Covid-19 pandemic. The teenagers told history and Latin teacher Claudia Marino about a locked basement door, leading her and colleagues to an abandoned boiler room and then to underground walls, frescoes, stucco and mosaics. The Special Superintendency of Rome began excavations in September 2025, revealing a well-preserved domus in an area once associated with elite figures such as Cicero, Pompey and Octavian. Archaeologists have recovered artifacts including amphorae, cups and lead-pipe inscriptions naming former occupants L. Fabius Gallus and Umbria Albina, and they believe the exposed rooms are only part of a larger site that could eventually open to students and visitors.

1d, 210p, 145 comments

Improvement in advanced Alzheimer’s disease following high-dose psilocybin

A case study published in Frontiers in Neuroscience documents significant but transient functional improvements in an octogenarian Japanese-American woman with advanced Alzheimer's disease following administration of high-dose psilocybin-containing mushrooms. After receiving 5g of psilocybin-containing mushrooms (Enigma strain), the patient, who had experienced 10 years of cognitive decline including 5 years of predominantly monosyllabic speech, spontaneously began autobiographical conversation approximately 19 hours post-administration. Over subsequent weeks, she demonstrated multidomain improvements including restoration of urinary continence after years of incontinence, improved ambulation, autonomous dressing, increased emotional responsiveness, sustained social interaction, contextual memory retrieval, and spontaneous conversational engagement. A second session with 3g a month later continued to show benefits. The researchers note these findings do not represent disease reversal but suggest that residual functional capacity may persist in advanced Alzheimer's disease and become temporarily accessible through psilocybin-induced modulation of large-scale brain networks.

3h, 16p, 11 comments

Paul Graham Is Strawmanning the Left's Argument Against Billionaires

Tony Onodi critiques Paul Graham’s Oxford Union speech defending billionaire wealth, arguing that Graham attacks a weak version of the left’s claim that “all billionaires are cheaters” while ignoring the stronger point that extreme wealth is inherently undeserved and concentrates power. Onodi notes Graham’s simplistic formula—growth rate and market size—as the basis for wealth, claiming it disproves the need for cheating, but says he fails to address deeper concerns about fairness, property rights, and societal impact, which he believes are more compelling arguments against billionaires.

23h, 90p, 34 comments

Show HN: machine0 – Persistent NixOS VMs You Control from the CLI

machine0.io offers on-demand, persistent virtual machines running NixOS or Ubuntu that can be managed entirely from the command line. Users receive dedicated CPU, RAM, optional GPUs and a static public IP with an HTTPS endpoint at .mac0.io, billed per minute starting at $0.013 hr⁻¹ and with the ability to suspend VMs to stop charges. The service supports reproducible builds via NixOS flakes or Ubuntu with Ansible, provides snapshot and image cloning, and promises 99.99 % uptime across four global regions. Installation is a single scripted command, and examples show creating development boxes, running autonomous agents, and hosting web services. Pricing scales with resource size, and unused credits are refundable.

6h, 43p, 7 comments

KDE Plasma 6.7 Released

Plasma 6.7 launches with major productivity upgrades, including the long‑awaited per‑screen virtual desktops, a quick light/dark theme toggle, and a Vietnamese lunar calendar. New usability tweaks streamline desktop tasks: faster virtual‑desktop switching, drag‑and‑drop favorites, improved Discover software center, and “type‑ahead” file selection. The release revives the classic Oxygen theme, enhances Breeze, and introduces the Union theming system that unifies styling for Plasma, QtQuick, and QtWidgets via CSS. Graphics and performance get boosts with simultaneous ICC color management and HDR support, lower CPU/GPU power use, and refined Wayland integration. Additional features cover network‑connection duplication, push‑to‑talk mute shortcuts, mixed‑skin‑tone emojis, enhanced system‑tray app monitoring, and richer printer handling. The update is dedicated to longtime KDE supporter Eric Laffoon.

1d, 590p, 229 comments

CrankGPT

CrankGPT is a local, human-powered AI solution designed for privacy and sustainability, operating entirely on-device to prevent data from being monetized by tech giants. It contrasts with cloud-based AI, which the text accuses of relying on gas-burning infrastructure and exacerbating climate change. The product offers tiered models—from basic "hand-cranked" devices for home use to high-power pedal-powered systems for enterprises—while critiquing wealth concentration by opposing tech CEOs' profit-driven models. By requiring physical effort to generate "tokens," it positions itself as an eco-friendly, cost-effective alternative to centralized AI services.

1d, 734p, 496 comments

What happened to nerds?

What the Fuck Happened to Nerds examines how the tech industry’s once‑trusted “nerd” image—embodied by figures like Steve Jobs and Steve Wozniak, who were seen as product‑obsessed perfectionists—has degenerated into a self‑promotional, attention‑seeking culture dominated by egocentric founders and venture‑capitalists. The author traces three phases: early founders who stayed behind the product, the rise of the founder‑as‑parable in the 2007‑2015 era, and the current period where leaders treat their influence as media content, exemplified by the controversial Founders Fund Mafia reality‑show video. Arguing that this shift erodes public trust, the piece urges tech leaders to reclaim authentic nerd values—curiosity, humility, and a focus on genuine product excellence—rather than chasing fame and shallow hype.

29m, 26p, 5 comments

Fable ban was never about a jailbreak

The U.S. Commerce Department issued an export‑control letter that forced Anthropic to shut down its flagship models, Fable 5 and Mythos 5, citing vague national‑security concerns rather than a proven jailbreak. Anthropic suspects the action stems from a reported guard‑rail bypass paper—an issue security experts say does not merit export restrictions—while critics argue the move is heavy‑handed, politically motivated, and sets a risky precedent for government control over American AI tools.

2h, 146p, 89 comments

Google Chrome update will close the door on ad blockers

Google Chrome’s upcoming update will finally eliminate support for Manifest V2 extensions, effectively ending the operation of popular ad blockers like uBlock Origin. A Chromium commit has removed the “kExtensionManifestV2Disabled” flag, which previously allowed legacy extensions to run, and Chrome 150 (expected June 30, 2026) will strip this workaround, with Chrome 151 (July 2026) removing the remaining MV2 flags. Google cites security risks and technical debt as reasons, while noting other Chromium‑based browsers could keep supporting MV2 if they choose. This move closes the last loophole for ad‑blocking extensions in Chrome.

6h, 388p, 222 comments

Feds freaked over Fable 5 after simple 'fix this code' prompt, not jailbreak

Fable 5 and Mythos 5 were placed under U.S. export controls after researchers asked the models to “fix this code,” a simple prompt that caused the AI to generate patches and test scripts for vulnerable code. Katie Moussouris of Luta Security, who read the underlying paper, says the request was a routine defensive task, not a jailbreak, and argues the restriction harms cybersecurity defenders. The Trump administration’s directive led Anthropic to disable the models for all customers, while over 100 security leaders signed an open letter urging the ban’s reversal, warning it will impede defenders while attackers gain an advantage.

3h, 58p, 32 comments

'Wow, it really worked ': 70s TV show causing worldwide panic today

Alternative 3, a 1977 Anglia Television mock‑documentary presented by Tim Brinton, claimed that governments were planning a secret moon‑to‑Mars colonisation scheme and that scientists were vanishing or dying to hide it. Though intended as a hoax, many viewers took it seriously, sparking panic and a flood of calls. The fiction inspired a book, entered American conspiracy circles, and was later cited by figures such as Milton William Cooper and Alex Jones, influencing UFO lore and pop culture. Today, revived claims of “missing scientists” linked to UFOs echo the original hoax, prompting even the Trump administration to launch an investigation, illustrating how a well‑crafted spoof can fuel enduring conspiracy theories.

1d, 347p, 54 comments

My Homelab AI Dev Platform

Rsgm describes how he built an AI‑enhanced homelab development platform using OpenCode web UI combined with GitOps on Arcane. He runs OpenCode on a VM with Git access, allowing the AI to generate code changes, push feature branches, and create pull requests that he reviews before merging. This workflow automates container updates, health‑check additions, and network configuration across dozens of Docker‑Compose stacks, cutting hours of manual work to minutes. While the system lacks CI feedback from Forgejo, it enables secure, device‑agnostic changes to home infrastructure without giving AI direct service access.

2h, 4p, 0 comments

Swish – Jupyter-like computational notebook for SWI-Prolog

SWISH is an online platform that provides a web‑based interface for SWI‑Prolog, allowing users to write, run, and share Prolog code directly in their browser. It supports real‑time collaboration features such as a shared editor, chat, and hangout sessions, enabling multiple users to work on the same file simultaneously. The service displays the number of users currently online and includes tools for managing sessions and logging out.

22h, 151p, 110 comments

Claude Corps

Claude Corps is a new national fellowship launched by Anthropic, funded with an initial $150 million, to place 1,000 early‑career fellows in U.S. nonprofits for a year‑long, full‑time, in‑person role. Fellows receive an $85,000 salary, benefits, extensive Claude AI training, and a large token budget, while host organizations—ranging from food banks and veteran support groups to coding education providers and marine conservation NGOs—gain AI expertise to boost their missions. The program, administered with CodePath and Social Finance, opens applications for the first 100‑person cohort in October 2026, with subsequent cohorts in early 2027, and aims to scale nationally and eventually internationally, measuring impact and open‑sourcing core technology.

6h, 10p, 1 comment

Hans Schulz – The father of the VEF Minox lens?

Hans R. Schulz – a little‑known German optics professor – was the engineer who calculated the impossibly tiny, high‑speed Minostigmat lens that gave the VEF Minox sub‑miniature camera its remarkable sharpness. In the mid‑1930s Walter Zapp’s mechanical design lacked a usable lens, and while major firms such as Agfa and Leitz dismissed the project, Schulz applied his expertise in physiological optics and lens theory to devise a 7 mm‑long, f/3.5 Cooke‑triplet that corrected astigmatism, field curvature and other aberrations without an adjustable aperture. Using painstaking hand calculations and glass selections from Schott, he produced a lens capable of resolving fine detail on the 8 × 11 mm negative, enabling the Minox to become a professional espionage tool and ensuring the camera’s commercial success for decades. Schulz died in 1968, his contribution largely unrecognised until recent archival research revealed his pivotal role.

5h, 9p, 3 comments

Europe Wants Digital Sovereignty. 2,165 Polish Organisations Show the Gap

Europe’s digital sovereignty drive faces a stark gap in Poland – CipherCue’s June 2026 scan of 2,165 Polish organisations (public‑sector, finance, healthcare and private firms) found that while 82% host their sites locally and the only major foreign host is France’s OVHcloud, the perimeter layer is dominated by US providers. All 128 detectable commercial CDNs were US‑headquartered (Cloudflare 89.8%, Amazon 7.0%, Akamai 2.3%, Fastly 0.8%) and every one of the 140 observed hosted‑mail services was Microsoft 365 or Google Workspace. Overall, 18.5% of organisations expose at least one US‑HQ component, with higher exposure in finance (19.7%) and healthcare (19.5%) than the public sector (10.6%). European alternatives are present only minimally, notably Cookiebot (Denmark) and Sophos (UK). The findings highlight that while Polish entities prefer European hosting, they remain reliant on US‑based CDN and email services, underscoring a key target area for European sovereign‑tech vendors.

3h, 13p, 0 comments

The White House Is Ratcheting Up Its War Against Anthropic

The article argues that the Trump administration’s chaotic and contradictory AI policy is undermining U.S. leadership in the field. While President Trump publicly pledged to accelerate AI development to beat China, the White House abruptly imposed export controls on Anthropic’s most advanced models, Fable 5 and Mythos 5, labeling them a national‑security threat and forcing the company to shut them down for all users. Critics, including cybersecurity experts and legal scholars, say the move is overly blunt, harms U.S. defenders, lacks clear justification, and reflects ideological bias against Anthropic. The episode highlights the administration’s erratic approach—simultaneously loosening chip exports to China, threatening a key domestic AI firm, and hinting at a de facto licensing regime—creating uncertainty for investors and innovators and casting doubt on Washington’s ability to steer America’s AI race effectively.

8h, 14p, 4 comments

Show HN: SharkClean MCP

sharkclean-mcp is an unofficial MCP server that lets any MCP client (e.g., Claude Code or Claude Desktop) control SharkClean/SharkNinja robot vacuums via SharkNinja’s current “skegox” cloud backend. It supports listing robots, checking status, starting whole‑home or room‑specific cleans, pausing, stopping, docking, setting power mode, and locating the robot, all through plain‑language commands. The project uses reverse‑engineered skegox API calls (originally from the shark2mqtt project) to work with newer models that have migrated away from the legacy Ayla IoT cloud. Setup requires Python ≥ 3.10, uv, and a one‑time browser login to obtain Auth0 tokens, after which credentials are cached. The server can be registered with MCP clients via configuration files, and a command‑line tool (shark-mcp-rc) offers direct scripting. The code is MIT‑licensed, includes unit tests, and notes that SharkNinja may change the backend at any time.

17h, 55p, 17 comments

An O(x)Caml book that runs

KC Sivaramakrishnan describes a new OCaml course on the NPTEL MOOC platform that is delivered as an interactive website rather than a PDF or static site. The pages embed client‑side OCaml cells powered by the x‑ocaml WebComponent (OCaml 5.4 compiled to JavaScript with Merlin, autocompletion and formatting) and, for advanced modules, a full 32‑bit Alpine Linux VM running in the browser via a WebAssembly x86 emulator. This “zero‑install” approach lets learners write, run, test, and even compile C or unikernels entirely in their browser, avoiding the typical setup hurdles. The content, generated from a single Markdown source, produces lecture pages, reveal.js slides, and runnable quizzes that are automatically checked in continuous integration to prevent code rot. The author used Claude Code and other LLMs for drafting, applying extensive pedagogical guidelines and iterative review. The course covers functional programming fundamentals, testing, memory safety, the experimental OxCaml language, and MirageOS unikernels, and remains open for free enrolment until 27 July 2026.

1d, 889p, 304 comments

Your ePub Is fine

André Klein recounts how his DRM‑free EPUB, which passed epubcheck 3.3 and worked on Kindle, Apple Books and Thorium, failed on Kobo devices, displaying a “corrupted” error. He discovered the cause was Kobo’s reliance on Adobe’s outdated RMSDK rendering engine, which cannot parse modern CSS 4 features such as max-width: min(150px, 30vw). Disabling the stylesheet or replacing the line with a simpler max-width: 150px allowed the book to load. Klein argues that epubcheck alone is insufficient for Kobo compatibility and urges publishers to test against Adobe’s engine, highlighting the broader issue of legacy software hindering open‑standard ebook publishing.

3h, 14p, 2 comments

SpaceX to acquire Cursor for $60B in stock, days after blockbuster IPO

SpaceX has agreed to acquire AI coding startup Cursor for $60 billion in stock, finalizing a deal announced days after SpaceX’s historic IPO. The acquisition aims to boost SpaceX’s AI division, built around Elon Musk’s xAI, which has faced controversies and restructuring. Cursor, founded in 2022 and previously valued around $29 billion, was poised for a $2 billion funding round before the deal. The purchase is expected to close in Q3 2026 and is intended to help SpaceX deliver on its massive AI infrastructure and enterprise application market ambitions.

1d, 115p, 105 comments

Ported my C game to WASM, here's every bug that I hit

Porting Match Morphosis to WebAssembly – The author describes the challenges of moving a C‑based game (using bgfx, SDL2, miniaudio, and cimgui) to the web with Emscripten. Key issues included needing a 32‑bit Visual Studio build to match WASM’s 32‑bit address space, fixing struct layout problems caused by 64‑bit pointers in serialized assets, debugging in native 32‑bit rather than the browser, correcting memory allocation errors that were hidden on 64‑bit, adapting to stricter OpenGL ES/WebGL rules (vertex layouts, component counts, Y‑axis flips), recompiling shaders for GLSL ES (using mix() instead of lerp() and proper float literals), and handling Web Audio autoplay restrictions plus an Emscripten export change requiring explicit HEAPF32 export. Despite the setbacks, the port was completed in a weekend, the game is live on itch.io, and the author praises the Emscripten toolchain once the platform‑specific pitfalls are addressed.

5h, 11p, 0 comments

Celebrating seven years of the Fairphone 3

Fairphone celebrates seven years of support for the Fairphone 3, announcing its retirement after the final software update in June and official end of support in August 2026 – two years beyond the originally planned five‑year lifespan. The article highlights the phone’s milestones, including carrier availability across Europe, enhanced modularity and recycled‑plastic content, the launch of the living‑wage bonus program that has paid out over $1.25 million to supply‑chain workers, and successful reuse of parts such as displays, motherboards and speaker modules. Fairphone urges owners to switch to alternative operating systems like LineageOS, /e/OS or postmarketOS to extend usability, and promotes its next model, the Fairphone Gen 6, which promises support until 2033, a five‑year warranty and up to 12 replaceable parts.

17h, 33p, 23 comments

When Did White-Collar Work Start to Look So Bleak?

The article traces how the promise of secure, fulfilling white‑collar work has unraveled from the 1980s yuppie boom to today’s precarious graduate labor market, weaving together reviews of Jodi Kantor’s “How to Start,” Dylan Gottlieb’s “Yuppies,” and Noam Scheiber’s “Mutiny.” It recounts how yuppie culture glorified long hours, fitness, and elite consumption while eroding genuine career advancement, leading to a generation of overworked associates and lawyers. The piece highlights contemporary graduate disillusionment, the rise of low‑pay, service‑industry jobs among degree holders, and growing union activism at places like Starbucks and Apple, while noting the looming threat of AI to creative fields. Ultimately, it argues that the illusion of control once offered by corporate work is fading, urging cross‑generational collaboration to confront a labor landscape defined by instability and anxiety.