The article addresses this point with the following:

> It would of course be possible to sic Chatty-Jeeps on the raw markup and have it extract all of this stuff automatically. But there are some good reasons why not. > > The first is that large language models (LLMs) routinely get stuff wrong. If you want bots to get it right, provide the metadata to ensure that they do. > > The second is that requiring an LLM to read the web is throughly disproportionate and exclusionary. Everyone parsing the web would need to be paying for pricy GPU time to parse out the meaning of the web. It would feel bizarre if "technological progress" meant that fat GPUs were required for computers to read web pages.