Published on June 15, 2026 at 06:00 CEST (UTC+2)
Your ePub Is Fine. Kobo Disagrees. Blame Adobe (275 points by sohkamyung)
Your ePub Is Fine. Kobo Disagrees. Blame Adobe
This article discusses issues where Kobo e-readers reject EPUB files that appear valid, tracing the problem back to Adobe’s rendering engine and metadata handling. The author demonstrates that the files conform to standards but fail due to subtle differences in how Adobe parses them. It highlights the frustration of compatibility problems in digital publishing.
Write for One Person (146 points by evakhoury)
Write for One Person
A comic by Wizard Zines that advocates writing content as if addressing a single, specific reader rather than a broad audience. The approach is meant to make communication more personal, engaging, and clear. It encourages creators to focus on one individual’s perspective to improve clarity and connection.
Even More Batteries Included with Emacs (24 points by signa11)
Even More Batteries Included with Emacs
This is the third in a series highlighting obscure but useful built-in features of the Emacs text editor. The author notes that many powerful Emacs capabilities are poorly documented or rarely discussed, and aims to surface them for veteran users. Features like undo-in-region and other lesser-known tools are showcased to improve productivity.
Show HN: Kage – Shadow any website to a single binary for offline viewing (451 points by tamnd)
Show HN: Kage – Shadow any website to a single binary for offline viewing
Kage is a tool that clones a website into a local folder for offline browsing by capturing the DOM via headless Chrome, stripping all JavaScript, and downloading associated assets (CSS, images, fonts). The result looks like the live site but runs no code. It’s useful for archiving, privacy, or reading content without network dependency.
21 years and counting of 'eight fallacies of distributed computing' (2025) (29 points by teleforce)
21 years and counting of 'eight fallacies of distributed computing' (2025)
This article revisits the classic list of mistaken assumptions about distributed systems (e.g., network is reliable, latency is zero). It traces the origins to Sun Microsystems engineers Bill Joy and Tom Lyon, and argues that despite decades of progress, developers still routinely violate these principles. The fallacies remain highly relevant for modern cloud and edge computing.
Bitsy (62 points by tosh)
Bitsy
Bitsy is a simple game engine designed for creating tiny, personal games, worlds, and stories. It emphasizes accessibility and creativity over technical complexity, providing a low-barrier entry for non-programmers. The site offers documentation, community forums, and links to games made with the engine.
Firewood Splitting Simulator (696 points by memalign)
Firewood Splitting Simulator
A playful web-based simulation where users drag a log and click to split it. It is a simple, interactive toy with minimal functionality, likely built as a fun distraction or demo of screen.toys. No AI or complex logic is involved.
Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model (302 points by unrvl22)
Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model
An issue on GitHub reveals that Rio-3.5-Open-397B, claimed by Rio de Janeiro’s technology agency IplanRIO as an original 397B model, is actually a weighted average (0.6 Nex / 0.4 Qwen) of two existing models. Evidence includes the model identifying itself as “Nex” and weight tensors matching the blend exactly. The discovery raises questions about transparency in public AI projects.
Show HN: Trace – Offline Mac meeting transcripts you can flag mid-call (119 points by AG342)
Show HN: Trace – Offline Mac meeting transcripts you can flag mid-call
Trace is a macOS app that captures microphone and system audio, transcribes locally on-device using Apple Silicon, and produces markdown transcripts. Users can flag important moments mid-call with ⌘K, and transcripts never leave the machine. It integrates with Notion, Obsidian, or AI chatbots for summarization, emphasizing privacy and offline operation.
A short history of Cerro Torre, the most controversial mountain (2012) (9 points by joebig)
A short history of Cerro Torre, the most controversial mountain (2012)
This article recounts the disputed first ascent of Cerro Torre in Patagonia, centered on Italian climber Cesare Maestri’s claims in 1959 and his controversial use of fixed bolts in 1970. More recently, climbers Hayden Kennedy and Jason Kruk removed those bolts, reigniting debate about ethics in mountaineering. The story illustrates long-standing controversies around achievement and fair means.
Local-first and privacy-preserving AI tools are gaining traction
Trace (offline meeting transcripts) and Kage (offline website archiving) both prioritize running entirely on-device, avoiding cloud uploads. This trend reflects growing demand for AI/ML solutions that respect user privacy, reduce latency, and work without constant internet. For developers, building local-first AI features (e.g., on-device speech recognition, small LLMs) can differentiate products in privacy-sensitive markets.
Transparency and provenance of LLMs are under scrutiny
The Rio LLM controversy shows that claimed “original” models can be simple merges of existing open-weight models, with no additional training. This highlights the need for verifiable model documentation, weight provenance tracking, and community auditing tools. AI/ML practitioners should expect increased pressure to publish training data, training logs, and reproducible build processes.
Distributed computing fallacies remain critical for AI infrastructure
The enduring relevance of the eight fallacies (e.g., network reliability, latency) directly applies to training large models across clusters and serving inference at scale. Ignoring these fallacies leads to failures in distributed training pipelines, inference batching, and edge deployment. Engineers must design for network failures, heterogeneous hardware, and dynamic topologies when building AI platforms.
Niche, non-AI tools still thrive alongside the AI boom
Bitsy and the Firewood Splitting Simulator demonstrate that simple, creative, or purely fun software continues to attract attention. The AI/ML ecosystem can learn from this: not every problem needs a large model, and small, focused tools (like small game engines or physical simulations) can deliver high user engagement without complex AI. There is value in building “delightful” products that solve specific human needs.
Discoverability remains a major UX challenge for powerful tools
The Emacs article’s theme—that powerful built-in features are often unknown to users—applies directly to AI/ML tools. Many developers are unaware of emerging features in frameworks (e.g., PyTorch, TensorFlow) or hardware capabilities (e.g., NPU acceleration). Improving documentation, tutorials, and in-app discovery can significantly reduce the learning curve and increase adoption of advanced AI techniques.
Meeting transcription with local LLMs is becoming practical
Trace’s offline transcription on Apple Silicon indicates that on-device speech recognition (using models like Whisper or custom ASR) can produce usable transcripts without cloud calls. This enables real-time note-taking in sensitive sectors (legal, medical, finance) where data cannot leave the device. The trend points toward wider deployment of small, efficient NLP models on consumer hardware.
Web archiving tools (like Kage) serve as data pipelines for AI training
While Kage strips JavaScript for offline viewing, its core function (cloning websites to static assets) can be repurposed for curating training datasets for vision-language models, layout understanding, or web navigation agents. The ability to capture rendered DOM ensures faithfulness to human-viewed content. Developers should consider such tools for building high-quality, structured web corpora.
Analysis generated by deepseek-reasoner