Published on July 24, 2026 at 06:00 CEST (UTC+2)
98.css (326 points by lopespm)
98.css
This is a CSS library for building user interfaces that faithfully recreate the look of Windows 98. It relies on semantic HTML and includes components like buttons, sliders, tree views, and progress indicators. The library has no JavaScript, making it compatible with any frontend framework. Accessibility is a primary goal, with required labels and aria attributes.
Writing by hand is good for your brain (1058 points by dwwoelfel)
Writing by hand is good for your brain
Author Neal Stephenson discusses research showing that handwriting engages more brain regions than typing. He explains the continuous micro-problem solving involved in spacing, letter connections, and arm coordination. He also notes that AI has pushed educators back to handwritten exams, creating challenges for students who never learned cursive and teachers who struggle with illegible writing.
The Corporate Creep of Plex: Why it may be time to move to Jellyfin (23 points by reprodev)
The Corporate Creep of Plex: Why it may be time to move to Jellyfin
This article contrasts Plex’s growing corporate control and feature bloat with the open-source alternative Jellyfin. It traces the author’s history from early XBMC to Plex’s convenience, then argues that Jellyfin now offers similar functionality without data collection or paywalls. The piece is the first in a “Thursday Thinkpiece” series.
The Visual 6502 (68 points by infiniteregrets)
The Visual 6502
This is an interactive JavaScript simulation of the MOS 6502 microprocessor, allowing users to explore its internal transistor-level operation. It uses HTML5 features and requires significant RAM. Controls include zoom, pan, and step-by-step simulation. Source code is available on GitHub, with links to additional emulators and assemblers.
Show HN: Echo – Fable-level results at 1/3 the cost using open-weight models (266 points by adam_rida)
Show HN: Echo – Fable-level results at 1/3 the cost using open-weight models
Echo is an AI system that combines multiple open-weight models (e.g., GLM-5.2, Kimi K2.7) instead of relying on a single model. It dynamically allocates compute and decides which models to use per request, achieving performance comparable to top proprietary models at much lower cost. The creator highlights how different models complement each other in surprising ways.
The Beam Engine (202 points by glinscott)
The Beam Engine
An interactive guide explaining how a beam engine works, from steam pressure to mechanical power. It builds the concepts step by step with 3D interactive figures—users can rotate, zoom, and see steam expansion, valves, and the beam’s motion. The engine produced about 15 horsepower, roughly the power of 150 people, and was a key driver of the Industrial Revolution.
Startup founders urge U.S. government not to shut off Chinese open weight AI (778 points by theanonymousone)
Startup founders urge U.S. government not to shut off Chinese open weight AI
A Politico article (content not available) reporting that a group of startup founders is lobbying the U.S. government to avoid blocking access to Chinese open-weight AI models. They argue such restrictions would harm innovation and increase costs for American startups that rely on these models.
What happened to TheNumbers.com (314 points by nickthegreek)
What happened to TheNumbers.com
TheNumbers.com, a trusted film industry data site, suddenly went down in March 2026 and returned weeks later with drastically reduced features. The author investigates and finds that an aggressive legal threat over its “authoritative” status—likely from a larger competitor—forced the owner to strip the site to avoid liability. The incident illustrates how fragile web data archives are.
What else do people draw on gradient.horse? (64 points by spython)
What else do people draw on gradient.horse?
A retrospective of drawings on gradient.horse, a site where users draw horses (and non-horses) using a gradient tool. Over 600,000 drawings exist, with common motifs including seahorses, skateboards, snails, crabs, philosophical text, and memes. The site uses an AI model to filter vandalism, but many creative non-horse drawings remain hidden unless the user enables “SHOW NON-HORSES.”
Why Software Factories Fail (or: harness engineering is not enough) (216 points by dhorthy)
Why Software Factories Fail (or: harness engineering is not enough)
This piece argues that attempts to build fully automated AI coding factories often fail because they focus too much on “loop engineering” (optimizing the prompt → code loop) and neglect context engineering, human oversight, and system design. The author’s company builds tools for human-agent collaboration, and he suggests that successful deployment requires balancing automation with human judgment.
Open-weight models as a strategic resource
The debate over restricting Chinese open-weight AI (article 7) and the success of Echo (article 5) highlight the growing importance of open-weight models. Founders argue that blocking access would increase costs and slow innovation, while projects like Echo show that combining multiple open-weight models can match proprietary performance at lower cost.
Why it matters: Open-weight models reduce dependence on a few providers, enable cost-efficient customization, and foster competitive diversity.
Implication: Companies should invest in model orchestration and multi-model strategies to hedge against policy risks and cost volatility.
Multi-model orchestration vs. single-model dominance
Echo (article 5) demonstrates a shift from using one LLM for everything to dynamically composing outputs from several models. This approach recovers performance that would otherwise require a larger, more expensive monolithic model.
Why it matters: It challenges the assumption that a single “best” model is optimal, and opens a path to more flexible, cost-effective AI systems.
Implication: AI teams should explore routing and ensemble techniques, and build evaluation frameworks that measure complementarity across models.
AI in education forces a return to analog methods
The rise of AI-written assignments is pushing educators back to handwritten, in-person exams (article 2). This creates practical challenges—students with poor handwriting, teachers unable to read it—and raises questions about assessment integrity in the AI era.
Why it matters: AI is reshaping educational practices beyond content generation, affecting testing, grading, and skill requirements.
Implication: Edtech solutions should develop hybrid approaches (e.g., AI-assisted grading of handwriting) rather than simply banning AI or reverting entirely to pen-and-paper.
Content moderation and AI’s hidden censorship
The gradient.horse site (article 9) uses an AI model to filter vandalism, but it also inadvertently hides many legitimate non-horse drawings. The default setting keeps them invisible, effectively censoring creativity.
Why it matters: This illustrates a common AI moderation trade-off: over-filtering to catch bad actors also suppresses benign content.
Implication: AI moderation systems need transparent, user-controllable settings (like “SHOW NON-HORSES”) and regular audits to reduce false positives.
Software factories require context engineering, not just loops
Article 10 argues that many AI coding factories fail because they focus on optimizing the code-generation loop while ignoring context management, human oversight, and error handling.
Why it matters: As AI coding agents move into production, reliable software engineering demands more than just prompt engineering—it requires system-level design for collaboration.
Implication: Teams building AI-assisted development tools should prioritize context injection, version control, and human-in-the-loop patterns over raw throughput.
The fragility of web data in the AI age
The sudden shutdown of TheNumbers.com (article 8) shows how a single legal threat can erase valuable datasets that many AI systems and research projects depend on. As AI training datasets increasingly rely on web-scraped data, this fragility poses a real risk.
Why it matters: AI models trained on such data may suffer from data availability bias or become outdated if sources disappear.
Implication: AI organizations should diversify data sources, archive critical datasets, and advocate for legal protections for non-commercial data curation.
Nostalgia and retro computing as a niche but persistent AI/UI trend
Projects like 98.css and Visual 6502 (articles 1 and 4) reflect ongoing interest in retro aesthetics and historical computing. While not directly AI, these projects influence UI design thinking and remind us that not every interface needs modern AI-generated features.
Why it matters: In an AI-saturated landscape, simplicity and familiarity can improve user trust and accessibility.
Implication: AI product designers should consider retro-inspired minimalism as a valid design pattern, especially for tools targeting older or privacy-conscious users.
Analysis generated by deepseek-reasoner