In contrast, new research that could be called "fundamental" in algorithms and theoretical computer science is typically several times longer and more complex. On one hand it seems intuitive that the more mature a field is, the higher the prerequisites for review and contribution. Computer science papers from the 60s seem much easier to read these days that what's been put out in the JACM since 2000, and modern papers in pure mathematics research are dense and incomprehensible (and commensurately more difficult to publish results in) even when placed next to modern research in computer science.
On the other hand, a fundamental improvement to Quicksort was developed in 2009 and the author's paper comes in only five pages . Is that because the improvement is "minor", because the author is exceptional at explaining the fundamental ideas clearly and succinctly, or because the author was relatively lazy about putting in the things usually demanded for publication (long sections on methodology, historical context, problem motivation, complexity analysis in different implementation contexts...)? It's hard to know whether some of these academic papers actually require their significant length, or if the presentation of material is simply inefficient. Cuckoo hash  was developed in 2001, but its paper comes in at 26 pages despite it being a fairly intuitive construction. The authors didn't just explain the fundamental result, they included two full pages of reference citations, a section on empirical performance characteristics and complete details of their experimental lab setup.
I'm not advocating that modern papers should necessarily be shorter, but I think it's an interesting question.
Gratuitous complexity builds over time. (And there is no subject area where I have seen such gratuitous complexity as with computers.)
I have always thought this was because either authors assume knowledge of the earlier papers and/or they have become "too familiar" with the concepts, but I could be wrong. In the later case, perhaps an analogy would be a singer who has to perform the same song thousands of times. When an artist sings it for the 999th time it may sound different and maybe there is assumption everyone has heard it before.
Whatever the reason, it is common to see the fundamental concepts either glossed over or not even mentioned in recent papers.
The exception to this, for me, where subsequent papers are helpful, is when they succinctly summarise prior research or track a subject area over time, e.g., annual reviews.
Anyway, when I am searching for older papers, it often feels "counterintuitive" since I see many people studying an area for the first time who want only the latest research. Perhaps an analogy here could be software. There seems to be a fascination on HN with software that is "new" and being changed on a daily basis versus software that is finished and has been in use for a long time.
A related idea is university textbooks. Many students believe they must have the latest editions, for a variety of reasons. However the changes from one edition to the next may be relatively small and easily summarized. IME, sometimes a topic may be more clearly explained in earlier editions. Reading the earlier edition's treatment of a topic followed by the later edition's treatment is sometimes illuminating in ways that reading only the latest edition is not.
In summary, I find it easier to read the old, 1-3 page paper that lucidly discloses an original concept and then build understanding from there, reading subsequent research, than to read something written, e.g., this year and assume that old research, by virtue of it date, is just a minor detail and no longer important.
For Cuckoo hashing, the original publication might be difficult to digest, but the journal version  improved it already quite a bit. There is a great explanation with an actual analysis by one of its original inventors available at . It usually takes a long time to find the best way of describing an algorithm and showing why it works.
Did you mean "The native sort implementation"? Also "lunch"?
We've come a long way. Python isn't even the fastest of languages
$ time python3.6 -c "import random; l = random.choices(range(1_000_000), k=2_000); l.sort()"