The Weird Week Edition

A researcher at a digital humanities institute published a taxonomy of recurring AI image generation artifacts — the extra fingers, the wrong shadows, the hair that blends into backgrounds, the text that hallucinates into decorative patterns. What makes it interesting rather than obvious is the proposed explanations for each artifact type, drawn from what we know about diffusion model training and the statistical properties of image datasets. Understanding why the errors happen is more interesting than cataloguing them, and this is the most coherent attempt at that I've seen.

The Pew Research Center released a study this week finding that 38% of web pages that existed in 2013 are no longer accessible. For pages from 2003, the figure is 72%. The study focuses on links in news articles and government documents, which presumably had more institutional maintenance than average. The actual link rot rate for personal pages, forums, and informal web content is almost certainly higher. The internet is more ephemeral than it presents itself, and we've built significant portions of our information infrastructure on the assumption that it isn't.

No JavaScript. No WebAssembly. A working, playable implementation of Snake using only HTML checkboxes and CSS transitions, exploiting the sibling combinator and the :checked pseudo-class to create state. The game is unplayable on mobile and not very good as a game, and none of that is the point. The point is that someone looked at the CSS specification and concluded that it was theoretically Turing-complete and then spent three months proving it in the most inconvenient way possible. The write-up is better than the game.

GNU FreeFont is an open-source typeface family that has been in continuous development since 2002, with contributors across dozens of countries adding glyphs for scripts and characters that commercial typefaces typically don't support. The latest release adds 300 new characters for minority scripts in Southeast Asia. It's not the best-looking typeface for body text, but as a record of distributed, unglamorous open-source collaboration around something as specific as typography, it's remarkable. The git commit history reads like a long argument about serifs conducted across two decades.

An analysis of peer review acceptance rates across major computer science conferences found that the top venues now have acceptance rates below 15%, with some below 10%. The analysis notes that this is a recent change — rates were above 25% a decade ago — and argues that the drop reflects volume rather than quality improvement. More submissions means more rejection for identical work. The implication, which the authors make explicitly, is that academic publication has become more like a lottery than a judgment process, and that the field is allocating talent accordingly.

The Video Game Music Archive at vgmusic.com has been running since 1996 and hosts MIDI transcriptions of music from thousands of games. It still gets new submissions. The archive is searchable, hand-maintained, and has no recommendation algorithm. Browsing it is a good way to spend forty-five minutes you meant to spend on something else. The MIDI files play correctly in a browser with no plugins. The interface hasn't changed in a decade. None of this is presented as a feature, but it functions as one.

Stellarium is open-source planetarium software with over ten million downloads and a web version that requires no installation. It is maintained primarily by one person, Fabien Chéreau, who has been working on it since 2001. The desktop version renders 600,000 stars. The project accepts donations but has no monetization, no ads, and no investors. Chéreau gave an interview this week about sustainable open-source maintenance that's worth reading alongside his quarter-century commit history. One person. Ten million downloads. Still going.