10 ms·
Yes HTML is too heavy and too expensive for LLM. We are working on a text-based format more suitable for AI.
by fbouvier 8mo ago
Yes HTML is too heavy and too expensive for LLM. We are working on a text-based format more suitable for AI.
- httpteapot 8mo agoWhat do you think of the DeepSeek OCR approach where they say that vision tokens might better compress a document than its pure text representation? https://news.ycombinator.com/item?id=45640594 https://news.ycombinator.com/item?id=45640594 I've spent some time feeding llm with scrapped web pages and I've found that retaining some style information (text size, visibility, decoration image content) is non trivial.
- fbouvier 8mo agoKeeping some kind of style information is definitely important to understand the semantics of the webpage.