Large Language Models are bad at producing (and can be hacked by) ASCII art ⋅ This Super Metroid FAQ is perfectly justified in a monospaced font (and has an ASCII illustration) ⋅ Whitespace accounts for twenty percent of English texts ⋅ The Knuth-Plass line-breaking algorithm assigns costs to breaking the line at different points and then tries to minimize them ⋅ “Underfull \hbox (badness 10000)” ⋅ Typographical rivers occur when spaces on adjacent lines are coincidentally aligned ⋅ The PAWS dataset is an adversarial test set for paraphrase detection, but T5_Paraphrase_Paws is trained on the same dataset to generate paraphrases ⋅ Gutenberg varied his type’s letterforms in order to create two uniform grey columns of text ⋅ Variable Fonts have parametrically defined variants ⋅ Hito Steyerl notes that something as arbitrary as a JPEG of a leopard can be used instead of random noise in Stable Diffusion models ⋅ Sussman attained enlightnment on thinking about a randomly wired neural network ⋅ “alternate black and white markings […] blend to form what appears to be a uniform grey half-tone” ⋅ “Thayer’s wildly erroneous assertion that the coloration of all animals was meant to conceal them from predators” ⋅ The leopard got his spots from the Ethiopian, Best Beloved, and is not likely to change them again
