GoKawiil
Tech News
A research team has proposed a technique for retrofitting pretrained subword-based language models so they operate directly on raw UTF-8 bytes rather than tokenized text. The approach builds on existing subword-trained models instead of training byte-level models from scratch, addressing a gap between prior academic claims and the lack of real-world adoption of byte-level LLMs.
A developer, prompted by a Twitter debate over JSON versus binary formats, built a custom binary schema language called jBin. The project explains how encoding data in binary rather than plain text or JSON can shrink payload size, reportedly by up to 80% compared to JSON, by using type and length markers instead of verbose text representations.
A writer recounts memorizing the first 100 digits of pi three years ago after reading Joshua Foer's 'Moonwalking with Einstein.' They used a person-action-object (PAO) system, a memorization method that pairs pairs of digits with vivid mental images of people performing actions on objects, building on the classical method of loci or memory palace technique.
A developer named Jay published a proposal called UTF-8000 aiming to remove the size limits of standard UTF-8 encoding, and reached out to Ken Thompson and other researchers behind earlier extension schemes for feedback. Ken Thompson replied with a technical observation about a zigzag signed variant requiring full code-unit reads for string comparison, while another correspondent, Thomas Eugene Bishop, pushed back on the proposal's framing, objecting to language calling rival schemes 'rejected' and disputing the claim that Larry Wall's UTF-8 extension is 'inextensible'.
Inspired by a YouTube video explaining matrix operations using flags, a developer built a compact encoding system that stores country flags in as little as 11 bits by breaking them into stripes, colors, and simple shared elements like stars or crosses. The system deliberately excludes coats of arms, state or city flags, and Nepal's non-rectangular flag to keep the encoding tractable, accepting that decoded flags will be recognizable but not pixel-perfect.