1kpapers.com: One Year of AI Research, Summarized for Four Dollars

⬅️ Back to Articles

Hassan El Mghari (@nutlope) posted a blunt launch note for 1kpapers.com: take the top 1,000 research papers of the last year, summarize them, visualize them. Fun fact in the same tweet: the whole summarization run cost about $4 on DeepSeek V4 Flash. The site is the atlas. The $4 is the punchline that makes people click.

  1. The product is a year-map, not a dump of PDFs. Browse by month (Aug 2025 through Aug 2026), by lab, by topic, or by “most trending / cited / starred.” Trending picks lean on citation impact, official-code adoption, recency, and field-wide significance. If you have been skimming arXiv titles and feeling behind, this is a better door than another open-tabs binge.

  2. The corpus was frozen on purpose. Methodology starts with 8,262 Hugging Face Daily Papers candidates, keeps verified frontier-lab work, dedupes by arXiv ID, checks version-pinned PDFs, and locks exactly 1,000 papers from August 4, 2025 through August 4, 2026. That window is the product. It is not “all AI forever.”

  3. They actually read the PDFs. The frozen set is 30,681 pages and 102.7 million extracted characters. Short papers go in one pass. Long ones use a deterministic 50,000-character map-reduce flow instead of quiet truncation. That detail matters more than the pretty covers. Silent cutoffs are how summary sites lie.

  4. The $4 claim holds up in their cost table, with receipts. Same text, same chunks, same prompt, reasoning off, prices frozen August 5, 2026: DeepSeek V4 Flash at $3.99 total ($0.004 per paper), GPT-5.6 Luna at $6.00, Claude Haiku 4.5 at $35.76. They say it outright: this is a cost comparison, not a factual-quality ranking. Inference only. Download, extraction, storage, and engineering time are off the ledger.

  5. You might want this to be a quality bake-off. It isn’t one. Cheap summaries can be wrong in confident ways. Use the atlas to find papers and skim orientation copy, then open the PDF for anything you will cite, ship, or bet money on. The site even frames itself as a popularity-weighted field guide, not an exhaustive history.

  6. The live atlas is a bit larger than the benchmark. They fold in 18 verified Together AI papers and metadata (citations, repos, labs, topics) for a 1,018-paper map. Lab pages for OpenAI, Anthropic, Meta AI, DeepMind, DeepSeek, and others make “what did this org publish that actually moved?” a one-click question.

The takeaway: Open 1kpapers.com, pick one month you skipped or one lab you follow, read three summaries, then open the PDFs that still feel load-bearing. Treat the $4 story as infrastructure news, not as permission to stop reading primary sources.

Related TMFNK Content

Crepi il lupo! 🐺