Small JSON responses that compress to <1k would fit in a single packet, so I don't see the advantage of going from "65 bytes with normal Zstandard compression, vs 28 bytes when using the past response as a dictionary - 57% smaller."
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.
The "aha" moment for me was that, without this dict, the user is going to always request a full download of the data. For instance, let's say the NYT published an article and you read it. Then an editors note is added to the article. When you go back to read the article, the data transfer is miniscule. Now that is an edge case, but imagine a website that allows comments.. twitter.. reddit.. small text based pages that at first seem incosequential until you think about how we use the web, millions of users, returning to pages over and over again.
For me, my mental model of this structure is a LUT(key/value pair) wrapped in a Version Control(hash).
Now i think your comment is correct if we were to add how many requests the webpage is recieving and how frequently changes are happening to said webpage. My blog would recieve no benefits from implementing this tech, and using napkin math, my blog would need 1000 days to break even. Microsofts' blog however... less than a day, in theory.
Reddit/NYT would have to publish their changes without changing the dictionary, meaning some portions would be largely absent from the dictionary and have worse compression than gzip. Probably fine for NYT, something like Reddit might actually have worse ratios than gzip in that case.
cough winrar cough
