When I joined Google in 2009, we were on the tail-end of a latency optimization kick that Larry had started in 2007. At the time, we had a budget of 20K gzipped for the entire search results page. I remember working on the visual redesign of 2010, where we had increased the page weight from 16K to 19K and there was much handwringing at the higher levels about how we were going to blow our
entire latency budget on one change.
We did some crazy stuff to squeeze everything in. We would literally count bytes on every change - one engineer wrote a tool that would run against your changelist demo server and output the difference in gzipped size of it. We used 'for(var i=0,e;e=arr[i++];) { ... }' as our default foreach loop because it was one character shorter than explicitly incrementing the loop counter. All HTML tags that could be left unterminated were, and all attributes that could be unquoted were. CSS classnames were manually named with 1-3 character abbreviations, with a dictionary elsewhere, to save on bytesize. I ran an experiment to see if we could use JQuery on the SRP (everything was done in raw vanilla JS), and the results were that it doubled the byte size and latency of the SRP, so that was a complete non-starter. At one point I had to do a CSS transition on an element that didn't exist in the HTML, because it was too heavy and so we had to pull it over via AJAX, so I had to do all sorts of crazy contortions to predict the height and position of revealed elements before the code for them actually existed on the client.
A lot of these convolutions should've been done by compiler, and indeed, a lot were moved to one when we got an HTML-aware templating language. But it gave me a real appreciation for how to write tight, efficient code under constraints - real engineering, not just slapping libraries together.
Alas, when I left the SRP was about 350K, which is atrocious. It looks like it's since been whittled down under 100K, but I still sometimes yearn for the era when Google loaded instantaneously.