Time is a pretty good filter.
HackerPapers collects research papers people keep bringing up on Hacker News. The aim is simple: find ideas that outlast the announcement cycle, and keep the conversations that made them interesting.
Three people. Three months.
A paper needs mentions from at least three distinct HN usernames across at least three different UTC calendar months. We count comments, not submitted story links. A comment must contain at least 100 characters of its author’s own text, excluding link text and bare URLs. Deleted and dead comments are excluded.
A score you can inspect.
The HN score is the number of distinct months with a recommendation, plus 0.5 for each distinct month with a strong recommendation. A month with a strong recommendation contributes 1.5. It is calculated before excerpts are capped. This is our derived score, not HN karma or a measure of scientific validity.
Each paper shows up to 20 original comments, prioritizing different commenters. Critical and neutral mentions remain part of the paper trail. Repeated links in one comment count once.
Rules first. Judgment still matters.
This edition uses keyword rules for stance and topics. Phrases such as “highly recommend” signal a strong recommendation; explicit criticism is checked first. Labels apply to the whole comment, so multi-paper comparisons, negation, sarcasm, and quoted speech can mislead the rules. Neutral is the default when a clear signal is absent.
Titles, authors, and abstracts come from arXiv, Crossref, or publisher citation metadata. Descriptions are excerpts from abstracts, not generated claims. Unresolved metadata stays out of the visible catalog. ArXiv versions are grouped together; explicit DOI matches merge corresponding editions. Other cross-source duplicates may remain until verified.
Human quality review is pending. This is an inspectable, rule-labeled edition. No claim of a 90% human-reviewed acceptance rate is made.
An archive, with an end date.
This collection was extracted from 49,324,111 records in 239 monthly Parquet shards of the open-index HN archive. The snapshot ends August 21, 2026. It is a historical collection, not a live feed. The archive’s incremental files were excluded because local checks found timezone and completeness problems.
Your reading, on your device.
Bookmarks, reading history, and appearance preferences are stored in your browser. There are no accounts and no cross-device sync. For You ranks unseen papers by shared topics with papers you opened. You can clear that history on the For You page. Clearing browser storage removes your list.
Built on generous conversations.
Inspired by Karpathy’s HackerTube and HackerBooks. Data is sourced from the open-index Hacker News archive. Every excerpt links to its author’s original comment. This is an independent project, unaffiliated with Hacker News, Y Combinator, or the paper authors.
Papers and PDFs stay with their original publishers. Dataset attribution and content-rights details are in the open data notes.