[cyanydeez]

The toy examples aren't the savings. Do the calculations with a json list of 100 objects and you'll find that compress increase more significantly.

So yeah, one time object return isn't impressive. Once those objects are in an array, then there's a much more remarkable compression.

While reading, I started wondering if we'll see an LLM constructor that'll take a API and some actual browser use and create a model that maximizes these types of message-centric compression.