4 ms·
AI Search Is a Disaster
- mdp2021 4y agoI found the article amusing after its drastic tones. One notable thing it only mentions is a reference to something remembered by Ted Chiang one week ago on the New Yorker (article which I wanted to submit but did not have the time to properly explore), the Xerox lossy compression format JBIG2, used ten years ago in photocopiers: such format used recurrent similar chunks found around in the image, and reconstructed the whole by places those pooled pieces in the page. But similarity was overly fuzzy and could not discriminate actual important detail: so the JBIG2 format took labels in architectural plans, such as 'length: 1.234', and replaced all similar occurrences in the page - ten meters, one tenth, 4.321 - with the "chosen" 'length: 1.234'. Cheers. Some are making parallels (as a broad analogy) between the management of data compression in some recent architectures and the "remarkable" (faulty versions of) JBIG2.