I have a smart RSS feed reader that pulls in all the CS papers from arxiv.org (as well as top comments from HN, articles from The Guardian, phys.org and many other sources) and runs them through a few classifiers including:
1. An active trained classifier based on my own preferences
2. A model that predicts how many votes a headline will get on HN
3. A model that predicts the comment/votes ratio a headline will get on HN if it hits it big
4. A model that predicts if a headline will get terminally flagged on HN
There is also a clustering system that makes sure I see a diversity of articles. For the CS abstracts it does a great job of inferring if I like the topic. It learned very fast that I like articles about how to build classifiers and about applied programming languages but I don't like most theoretical CS articles. It is not so good at estimating the quality of work, although I think I could do better if I downloaded the PDFs because I could get the page count. My impression is that 2/3 of CS papers are not long enough to stand on their own, they might be OK to support a conference talk and maybe somebody who is working in the field every day but for an applications dev who wrote 1 PNAS paper on text classification a real long time ago I think papers under 5 pages are difficult to read without gathering another 50 pages worth of references.
I am not getting a sample of the "most important" work (might need to wait for papers to get some citations to access quality that way) and not reading many papers for deep comprehension, but I am seeing a lot of abstracts for papers that apply popular methods to a range of problems which is great for a practitioner.
On a good day the system ingests 1000 feed items (some of which are scientific papers) and wants to show me 300, I skim maybe 200. It tries to show me more articles that I "like" but if I "love" an article it gets on my favorite list and can go into further workflows.
(The system learns topicality of CS papers and some other topics really well but boy does it struggle learning that I like the NFL and don't like the Premier League. This weekend I may try using an embedding-based model to see if it does much better.)
I get recommendations in scholar.google.com based on my profile and sometimes browse arxiv.org to see the latest submissions. Also, I can say that https://www.quantamagazine.org/ is very nice. Other than that, I just follow my favorite scientists through their personal blogs or social media.
Twitter was good for this before Elon fucked it up.
I had lists of scientists, who'd share their work and that of others.
1) You could start with a good paper and download all its references. Than skim those and download their references.
2) Bulk download all papers (100 TB, 4 weeks, Sci-Hub, preferably in a country where this is legal) and search all those with a local search like Apple Spotlight.
3) Go through HN comments and download its linked papers https://news.ycombinator.com/threads?id=morphle
4) Follow Alan Kay, Richard Feynman, Carl Sagan, Stephen J. Gould, The Royal Institution, Stanford EE380, Nasa, Sabrina Hossenfelder, James Mickens and scrape/download their recommendations
5) Start a startup that automates this for everyone. I have written most of the scripts