Published on August 18, 2026 at 18:06 CEST (UTC+2)
Using the railway network as a flatbed scanner (199 points by otherayden)
Using the railway network as a flatbed scanner — The author describes a creative hardware project using an industrial linear scanning camera to capture extremely wide, high-resolution photographs from trains and ferries. By continuously capturing thin vertical lines as the vehicle moves and stitching them together, they produce panoramic images (like a 56,894x2,048 pixel grayscale image of the San Francisco-Oakland ferry route). The post traces the history of this technique back to 1990s digital scanning backs for large-format film photography, and includes a talk delivered at EMFcamp 2026. It's a technical deep-dive into the mechanics, challenges, and artistic possibilities of line-scan imaging.
Fixing a Bricked Framework Laptop (135 points by jp_sc)
Fixing a Bricked Framework Laptop — A Framework 13 laptop owner describes how a BIOS update (version 3.20, pushed via official newsletter) hung mid-flash and permanently bricked his AMD 7040-series machine. Framework support told him to drain the battery and reboot, but the laptop never recovered, and since the warranty had expired, he was told to buy a new motherboard for at least CA$500. He discovered many other users had experienced identical issues with that BIOS update and with the model's flashing process generally since March 2025. Instead of giving up, he documents how he successfully repaired the laptop himself using only about $20 worth of tools, highlighting broader issues with repairability despite Framework's marketing. The piece raises serious concerns about manufacturer accountability when vendor-pushed updates brick devices.
The Amazon Tax (309 points by herbertl)
The Amazon Tax — Seth Godin critiques Amazon's search advertising model, calling it "legal theft" rather than a true tax because it produces no public benefit. He notes Amazon makes nearly a billion dollars per week in search ad revenue — enough to give every employee a $35,000 bonus and still have money left over. Even his own publisher had to buy Amazon search ads to help people discover his new book, which means the marketplace is essentially monetizing the desperation of merchants and authors. The post argues these ads distort what buyers see and extract money from the very sellers who make Amazon valuable in the first place. It's a sharp commentary on platform power, marketplace dynamics, and the hidden costs of search-based commerce.
Linux 7.3 improves performance when running out of vRAM (387 points by flaburgan)
Linux 7.3 improves performance when running out of vRAM — This technical blog post (part of a series) celebrates kernel patches from the author finally being merged upstream for Linux 7.3, improving VRAM overcommit behavior in GPU drivers. The author explains that running out of VRAM should theoretically be a performance issue, not a stability one, since GPU drivers can overcommit VRAM — but in practice games crash and performance plummets. The post digs into why running out of VRAM historically sucks so badly and how the new kernel work makes it "suck as little as possible." It's a deep technical dive into GPU memory management, aimed at kernel/driver developers and people interested in Linux gaming performance. The previous post focused on stability for games within physical VRAM limits; this one tackles the harder question of exceeding those limits.
Python Polars Cheatsheet (based on our O'Reilly book) (61 points by jeroenjanssens)
Python Polars Cheatsheet — This is a reference resource (also available as PDF, and tied to the authors' O'Reilly book) for using the Polars DataFrame library in Python. It covers installation, basic imports, data structures (Series, DataFrame, LazyFrame), and common query patterns using a chain of method calls — like reading CSV, filtering, and writing parquet. The cheatsheet is designed as a quick visual guide for transforming, analyzing, and visualizing data with Polars. It highlights Polars' expressive API and fast performance as an alternative to pandas.
Teaching my kid to code with a modern MUD (126 points by andrewjanke)
Teaching my kid to code with a modern MUD — The author built a modern, web-based multiplayer text game called "Canon" to teach his eight-year-old daughter programming. The game is intentionally modeled on classic MUDs/MUCKs from the old internet, but redesigned so that every room, object, and interaction is player-created and can be inspected, copied, and modified. It includes a tiny scripting language called "Cant" that is deliberately simple and not a "good" language, as a learning tool. The goal is to give his daughter what "View Source" gave his generation: the ability to take anything apart and see how it works. The piece is part nostalgic tribute, part design essay on learning environments.
How I Under-Engineered My Book (15 points by goostavos)
How I Under-Engineered My Book — A short, humorous counterpoint to a more elaborate "How I Over-Engineered My Book" article. The author describes a deliberately simple workflow: typing first drafts into Obsidian, copy-pasting into Word, then dragging and dropping into Box for Manning Publications. That's it — no build systems, no custom tooling, no automation. The piece is essentially a self-deprecating, low-key celebration of doing the simplest thing that works. It's a reminder that not every creative workflow needs engineering.
One Oakland police officer made $490k in overtime (55 points by speckx)
One Oakland police officer made $490k in overtime — A data-driven investigative piece (partnership between The Oaklandside and UC Berkeley's Investigative Reporting Program) into Oakland Police Department's runaway overtime spending. The department's $386 million budget is roughly 19% of city spending, and overtime hit over $55 million last fiscal year — $31 million over budget. One officer made $490,000 in a single year, raising questions about "superhuman hours" and sparse record-keeping. The police union blames severe understaffing, while the city faces ~$120 million annual budget gaps through 2030. The reporting connects excessive overtime to attrition, creating a vicious cycle where understaffing drives overtime, which drives more officers out.
Fairphone is now officially available in the United States (184 points by Vinnl)
Fairphone is now officially available in the United States — Fairphone announces the Gen. 6+ model and its expanded availability in the US, shipping to every state except Alaska and Hawaii with local pricing and support. The company, now nearly 16 years old, emphasizes its core mission: proving smartphones can be made differently — lasting longer, repairable, and ethically sourced. The "plus" version focuses on upgrades that don't compromise repairability: more performance, more memory, and more personalization. The piece signals a strategic push into the American market for the ethical electronics company, at a time when right-to-repair is a growing policy conversation.
Babies born under sugar rationing grew into adults with lower cancer risk (67 points by zeristor)
Babies born under sugar rationing grew into adults with lower cancer risk — A study leveraged a natural experiment: post-WWII Britain kept sugar rationing until September 1953, then consumption nearly doubled. Researchers compared children born just before vs. just after rationing ended, looking at health outcomes into their 70s. Those with less sugar exposure in the first few years of life had lower risk of several cancers and showed signs of slower biological aging. They also continued consuming less sugar and had healthier diets as adults. The piece discusses the scientific value of these "natural experiments" when randomized controlled trials on babies are obviously impossible.
Edge AI meets niche hardware hacks — sensor fusion and unconventional data capture are becoming accessible. The railway-scanner project demonstrates that specialized capture hardware (industrial linear cameras) is becoming cheap enough for hobbyists, and the resulting very-high-resolution imagery is exactly the kind of data that feeds computer vision and multimodal model training. Trend implications: as edge hardware becomes affordable, the bottleneck shifts from sensor availability to data processing and model ingestion pipelines — expect more novel, high-fidelity datasets to emerge from unusual sources (trains, drones, satellites), demanding new preprocessing and stitching architectures in ML pipelines.
The "AI tax" on commerce is a growing friction point — and a market signal for AI-native discovery. Seth Godin's critique of Amazon's search-ad revenue (~$1B/week) highlights an important dynamic: as search and discovery get monetized to the point of distortion, users and sellers alike are increasingly open to alternatives. This matters for AI/ML because LLM-based shopping assistants (Perplexity, ChatGPT, personalized recommendation engines) are explicitly positioned as "unbiased discovery" alternatives to ad-driven search. Actionable takeaway: AI teams building consumer-facing recommendation or shopping products should position against ad-influenced discovery as a core value proposition — the perceived "tax" is a competitive opportunity.
GPU memory management is becoming a first-class AI infrastructure concern — kernel-level work matters. The Linux 7.3 vRAM overcommit work signals that as AI workloads (and AI-adjacent gaming/graphics workloads) push beyond physical GPU memory, software-level memory management is increasingly the differentiator. This matters for AI model deployment: inference and fine-tuning of large models frequently exceed single-GPU memory, and solutions relying on overcommit, swap, and careful scheduling are becoming production-critical. Actionable insight: teams should track kernel/driver-level memory management improvements, as they can yield significant performance stability gains for LLM serving without any model changes — and consider frameworks (vLLM, TensorRT-LLM) that implement similar logic at the application layer.
Dataset efficiency and "data as artifact" are replacing model-centric hype — The Polars cheatsheet, while simple, reflects a longer trend: the data-processing layer (fast DataFrame libraries, Arrow, Parquet, lazy evaluation) keeps getting more investment and mindshare even as model APIs commoditize. Trend implication: AI/ML engineers are realizing that for many products, the moat isn't the model but the data pipelines that feed it, and tools that make data processing faster, more memory-efficient, and more ergonomic (Polars vs. pandas) directly reduce AI product iteration time. The takeaway for teams: invest in data infrastructure and modern tooling as heavily as model experimentation.
Natural experiments and longitudinal data are providing more rigorous evidence for health-AI features, and "early-life factors + multi-decade outcomes" datasets are gold. The sugar-rationing study shows the analytical value of natural experiments that link early-life conditions to decades-later cancer and aging outcomes. For AI/ML, this points to growing opportunities for predictive health models trained on rich historical cohort data — and a caution about using retrospective observational data without accounting for confounders. As consumer health AI (aging clocks, cancer risk prediction, dietary recommendations) grows, the strongest products will be built on well-understood natural experiments and longitudinal cohorts, not just cross-sectional correlations.
The "under-engineering" backlash is a healthy counter-signal to AI over-automation. The deliberately minimal workflow article — and its implicit contrast with over-engineered book production pipelines — reflects a broader mood in the developer community. As AI coding tools tempt every team to build increasingly complex automation and agentic pipelines, there's a growing counter-movement toward simplicity, human judgment, and doing the direct thing. For AI/ML teams, this is a useful guardrail: not every workflow needs a copilot, an agent, or a pipeline — and the highest-leverage application of AI is often in the few places where it removes genuine pain, rather than automating things that were already fine.
Repairability, ethics, and accountability are becoming product differentiators in hardware — and data/AI governance will follow. Framework's bricking debacle and Fairphone's US expansion both point to a market where consumers are increasingly aware of planned obsolescence, repairability, and vendor accountability. For AI/ML, the parallel is clear: as AI systems get integrated into hardware (BIOS updates, on-device models, embedded agents), the "bricking" equivalent — silent model updates that degrade performance, privacy-invasive on-device AI, or vendor lock-in for ML features — will become a consumer trust issue. Companies building on-device AI should treat update safety, rollback, and user control as core product features, not afterthoughts — and the right-to-repair movement will increasingly extend to "right to understand and control the AI in your device."
Analysis generated by deepseek-reasoner