@randomgeek@sjn@ovid Autocomplete is just generation loss with some hash tables as guardrails. They’re both Markov processes. Therefore, LLM == copy machine.
@mjgardner I tend to adopt domain jargon pretty quick. Related to neurodivergent masking. I got used to terms having different effective definitions based on context.
@randomgeek@mjgardner I remember an issue like that where a range of scanners would replace digits by other digits in scanned documents and blueprints. It was a few years back. Might've been compression related.
They used the #JBIG2 image format’s lossy “pattern matching & substitution” method that substitutes previously-encoded characters if they look enough like the one currently being encoded.
This is a great analogy to how #LLM-based “#AI” works.
@mjgardner@randomgeek@sjn@ovid What is this symbol that keeps popping up between every syllable on the website? As a screen reader user, it makes it difficult to read the page.
@mjgardner I’m using NVDA on Windows. I just added a speech dictionary entry to ignore the soft hyphen. If you want to know how it sounded for me, try to type some text, but sep ar ate each sy li ble with a space like I just did, then use Voice Over to read it back.
@mjgardner Thanks for bringing that to my attention. It's rather annoying for those who simply wish to read a website with those characters and not write one IMO.
Unfortunately, there are many actors here: site authors, web publishing software #developers, #browser developers, #ScreenReader developers, and tying it all together @wai and the #accessibility standards they publish and support.
Mostly everyone keeps up with #a11y standards now, but sadly you’ve run into a long-standing gap with #NVDA and soft hyphen #HTML entities.
Add comment